Real-time prediction method and system for shield tunnel construction geology based on neural network

Abstract

The application discloses a shield tunnel construction geological real-time prediction method and system based on a neural network, and the method comprises the following steps: acquiring shield tunneling parameter data and preprocessing the shield tunneling parameter data, inputting the preprocessed shield tunneling parameter data into a pre-trained geological feature extraction model, and outputting geological feature data; then inputting the geological feature data into a pre-trained geological feature classification model, and outputting a geological category; the geological feature extraction model is a neural network model composed of an integrated self-encoder with an attention mechanism, and the integrated self-encoder comprises a plurality of parallel sub-self-encoders; and the geological feature classification model is a residual CNN neural network model. The application can fully utilize the tunneling parameters of the whole construction section to realize real-time and accurate geological prediction, reduces the total sample size of the required data for training, greatly improves the accuracy and timeliness of model prediction, and can more effectively distinguish the geological change condition.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of tunnel construction, and relates to a tunnel construction geological real-time prediction method and system, in particular to a shield tunnel construction geological real-time prediction method and system based on a neural network. BACKGROUND

[0002] With the rapid development of tunnel traffic, urban tunnel construction projects are gradually increasing. Shield tunnel construction gradually becomes the main method of urban underground tunnel construction due to its small vibration, high operation efficiency and less environmental impact. However, in the process of shield construction, due to the uncertainty of the geological conditions of the working face, the shield machine operator cannot reasonably adjust the tunneling parameters of the shield machine. Improper operation parameters make the shield machine unable to adapt to the changes in geological conditions well, resulting in abnormal tunneling parameters of the shield machine and abnormal load conditions. Further, the shield machine cutter is severely worn, the tunneling speed is slowed down, and the ground surface is abnormally settled, which seriously affects the construction safety and construction progress. In order to improve the construction efficiency and ensure the safety of construction and the surrounding environment, it is particularly important to accurately predict the geological information of the working face.

[0003] After searching, the invention of Chinese patent No. CN115659842A discloses a geological dynamic inversion method using a shield tunneling database, especially a large-diameter shield and a shield for constructing an underwater tunnel or a soft-hard stratum construction. The stratum of the large-diameter shield excavation face is complex, the geological survey of the underwater tunnel is not accurate, and the determination of the soil-rock interface of the soft-hard stratum is difficult. The project's geological survey data, tunnel longitudinal section diagram, etc. cannot accurately reflect the real geological survey situation. Through the collection and analysis of shield equipment data, spoil information and a plurality of sensor information, a data-driven stratum inversion is formed to continuously generate new electronic maps in shield tunneling to guide the development of shield construction, etc. The invention of Chinese patent No. CN108846521A proposes a shield construction adverse geological type prediction method based on Xgboost. The implementation steps are: preprocessing the PDV historical data of the shield machine; obtaining the key features of the preprocessed plurality of tunneling parameter data; constructing an adverse geological prediction data package; establishing an Xgboost algorithm adverse geological prediction model; evaluating the Xgboost algorithm adverse geological prediction model; and predicting the adverse geological type in the shield construction process.

[0004] However, the above-mentioned shield geological prediction related technology does not fully consider the incompleteness of the geological survey report, so that the tunneling parameter data cannot be fully utilized, and the geological prediction accuracy and timeliness are low.

[0005] The existing shield construction geological prediction related technology only uses the tunneling parameter data of the geological drilling section, does not fully consider the incompleteness of the geological survey report, so that the tunneling parameter data of the non-geological drilling section cannot be fully utilized. When there is a local geological mutation in the adjacent geological drilling section, the existing technology cannot accurately predict the geological type according to the change of the tunneling parameter, resulting in low accuracy and timeliness of the geological prediction. In addition, a large amount of tunneling parameter data cannot be fully utilized, so that the existing technology needs a larger total sample size when training the network.

[0006] Through retrieval, the invention of Chinese patent No. CN113431635A discloses a semi-supervised shield tunnel face geological type prediction method and system, which includes the following steps: screening machine parameters related to geological conditions from machine data; preprocessing the screened machine parameters; constructing a no-label data set and a labeled data set according to the preprocessed machine parameters; establishing a semi-supervised framework for predicting the geological conditions of the tunnel face, obtaining a geological feature extractor and a feature classifier; training the geological feature extractor using a constrained DenseNet auto-encoding network and a no-label data set; training the geological feature classifier using a deep neural network and a labeled data set, and finally realizing the prediction of the geological type in front of the tunnel face.

[0007] However, the existing technology cannot consider the multi-dimensional characteristics of the feature data, and a single feature extraction model cannot accurately extract multi-dimensional characteristics such as time and space, resulting in low prediction accuracy of the model.

[0008] Therefore, there is an urgent need for a method that fully utilizes the tunneling parameter data of the entire construction section for geological prediction and identification to reduce the total sample size of the training data, improve the accuracy and timeliness of the geological prediction during shield construction, and provide guidance for the active parameter setting of the shield machine in a timely manner. SUMMARY

[0009] To solve the above problems, the present application proposes a real-time geological prediction method and system for shield tunnel construction based on neural network. The present application fully utilizes the shield machine tunneling parameter to extract geological feature information for geological prediction, thereby providing guidance for shield machine operation parameter setting, improving construction progress and reducing construction risk.

[0010] The technical scheme adopted by the present application is as follows:

[0011] A real-time geological prediction method for shield tunnel construction based on neural network, comprising the following steps:

[0012] The shield tunneling parameter data is acquired and preprocessed, the preprocessed shield tunneling parameter data is input into a pre-trained geological feature extraction model, and geological feature data is output; then the geological feature data is input into a pre-trained geological feature classification model, and a geological category is output.

[0013] The geological feature extraction model is a neural network model composed of an integrated self-encoder with an attention mechanism, and the integrated self-encoder includes multiple parallel sub-self-encoders, each of which includes an encoder and a decoder; and the geological feature classification model is a residual CNN neural network model.

[0014] Further, the shield tunneling parameter data includes slurry pressure, pushing speed, pushing force, cutterhead torque, cutterhead rotating speed and synchronous grouting pressure; and the construction progress information includes tunneling time and tunneling ring number.

[0015] Further, the preprocessing of the shield tunneling parameter data specifically includes: performing abnormal value detection and elimination by box plot method; performing data denoising by wavelet transform; then matching the tunneling time with the shield segment ring spacing, replacing the original data with the average value and the standard deviation of the data within one ring spacing to realize information compression; and finally performing maximum and minimum normalization processing.

[0016] Further, the three parallel sub-self-encoders are respectively:

[0017] The first sub-self-encoder based on a CNN network is used for extracting intra-ring and inter-ring spatial features;

[0018] The second sub-self-encoder based on an LSTM network is used for extracting tunneling parameter time sequence features;

[0019] The third sub-self-encoder based on a CNN-LSTM hybrid network is used for extracting global features.

[0020] Further, the integrated self-encoder further includes an attention fusion layer, and the three features are fused by the following steps:

[0021] 1) The encoding features H1, H2 and H3 output by each sub-self-encoder are spliced into a feature matrix H;

[0022] 2) For each encoding feature, an attention score is calculated by a LeakyReLU activation function;

[0023] 3) The attention score is subjected to a global average pooling operation;

[0024] 4) The pooled attention score is subjected to a Softmax function to generate normalized weights, so that each weight value is in the interval [0, 1] and the sum is 1;

[0025] 5) weighted sum according to the normalized weight of each coding feature, output the fusion feature.

[0026] Further, the pre-training method of the geological feature extraction model and the geological feature classification model is:

[0027] Obtain shield tunneling parameter data, construction progress information and geological survey report information;

[0028] Preprocess the shield tunneling parameter data, and cluster the geological types according to the soil physical properties; construct the shield tunneling parameter data set from the shield tunneling parameter data and the construction progress information; link the preprocessed shield tunneling parameter data and the geological types according to the construction progress information, and construct the shield tunneling geological data set;

[0029] Divide the shield tunneling geological data set into unsupervised learning data set and supervised learning data set, train the geological feature extraction model using the unsupervised learning data set, and train each sub-auto-encoder independently; then freeze the parameters of the geological feature extraction model, and train the geological feature classification model using the supervised learning data set.

[0030] Further, the clustering of the geological types according to the soil physical properties specifically includes: based on the geological survey report information, taking the internal friction angle, cohesion and elastic modulus of each type of soil in the project as the clustering analysis features, performing K-means++ unsupervised clustering analysis, and dividing the soil into several geological types.

[0031] Further, the unsupervised learning data set is a shield tunneling parameter data set without geological type label, and the supervised learning data set is a shield tunneling geological data set with geological type label.

[0032] Further, in the training process of the geological feature classification model, the shield tunneling parameter data in the supervised learning data set is first input into the trained geological feature extraction model for feature extraction, the output geological feature data is matched with the labeled geological type label, and then the geological feature classification model is trained.

[0033] A shield tunnel construction geological real-time prediction system based on neural network, for the above method, comprising:

[0034] Data acquisition module: for acquiring shield tunneling parameter data and preprocessing;

[0035] Feature extraction module: for inputting the preprocessed shield tunneling parameter data into the pre-trained geological feature extraction model, and outputting geological feature data;

[0036] The feature classification module is used for inputting the geological feature data output by the geological feature extraction model into a pre-trained geological feature classification model, and outputting a geological category.

[0037] The current shield construction geological prediction technology only extracts features from multi-dimensional data through a single feature extraction model, and the effectiveness of feature extraction is low, and the prediction model accuracy is insufficient. Compared with the prior art, the geological real-time prediction method proposed in the present application can extract multi-dimensional features such as time and space of input data based on different function feature extraction models, greatly improving the accuracy and timeliness of model prediction, and more effectively distinguishing geological changes. Thus, it provides guidance for shield operators to set shield active parameters. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The figure is a schematic diagram of the geological feature extraction model based on the integrated autoencoder with attention mechanism in the embodiment of the present application.

[0039] Figure 2 The figure is a schematic diagram of the geological feature classification model based on the convolutional residual neural network in the embodiment of the present application. DETAILED DESCRIPTION

[0040] The technical solutions of the present application will be further clarified, detailed and described below in combination with the drawings and specific embodiments.

[0041] A shield tunnel construction geological real-time prediction method based on a neural network, comprising the following steps:

[0042] Obtain shield machine tunneling parameter data and pre-process the data, input the pre-processed shield machine tunneling parameter data into a pre-trained geological feature extraction model, and output geological feature data; then input the geological feature data into a pre-trained geological feature classification model, and output a geological category.

[0043] The shield machine tunneling parameter data includes mud pressure, pushing speed, pushing force, cutterhead torque, cutterhead speed and synchronous grouting pressure; the construction progress information includes tunneling time and tunneling ring number. The pre-processing of the shield machine tunneling parameter data specifically includes: abnormal value detection and elimination by box plot method; data denoising by wavelet transform; then matching the tunneling time with the shield machine segment ring spacing, replacing the original data with the average value and standard deviation of the data within one ring spacing to realize information compression; finally, performing maximum and minimum normalization.

[0044] The geological feature extraction model is a neural network model composed of an integrated autoencoder with an attention mechanism, the integrated autoencoder comprising a plurality of parallel sub-autoencoders, each sub-autoencoder comprising an encoder and a decoder; and the geological feature classification model is a residual CNN neural network model.

[0045] The three parallel sub-autoencoders are: a first sub-autoencoder based on a CNN network, used to extract intra- and inter-ring spatial features; a second sub-autoencoder based on an LSTM network, used to extract time series features of the tunneling parameters; and a third sub-autoencoder based on a CNN-LSTM hybrid network, used to extract global features.

[0046] The integrated autoencoder further comprises an attention fusion layer, which fuses the three features by the following steps:

[0047] 1) concatenating the encoding features H1, H2 and H3 output by each sub-autoencoder into a feature matrix H;

[0048] 2) for each encoding feature, calculating an attention score through a LeakyReLU activation function;

[0049] 3) performing a global average pooling operation on the attention scores;

[0050] 4) generating normalized weights from the pooled attention scores through a Softmax function, so that each weight value is in the interval [0, 1] and the sum is 1;

[0051] 5) performing weighted summation according to the normalized weights of each encoding feature to output the fused feature.

[0052] The pre-training method of the geological feature extraction model and the geological feature classification model is as follows:

[0053] Obtaining shield tunneling parameter data, construction progress information and geological survey report information;

[0054] Preprocessing the shield tunneling parameter data, clustering the geological types according to the soil physical properties; constructing a shield tunneling parameter dataset from the shield tunneling parameter data and the construction progress information; linking the preprocessed shield tunneling parameter data and the geological types according to the construction progress information to construct a shield tunneling geological dataset;

[0055] Dividing the shield tunneling geological dataset into an unsupervised learning dataset and a supervised learning dataset, training the geological feature extraction model using the unsupervised learning dataset, and training each sub-autoencoder independently; then freezing the parameters of the geological feature extraction model, and training the geological feature classification model using the supervised learning dataset.

[0056] The clustering of the geological types according to the soil physical properties specifically comprises: taking the internal friction angle, cohesion and elastic modulus of each type of soil in the project as the characteristics of the clustering analysis based on the geological survey report information, performing K-means++ unsupervised clustering analysis, and dividing the soil into several geological types.

[0057] The unsupervised learning data set is a shield tunneling parameter data set without a geological type label, and the supervised learning data set is a shield tunneling geological data set with a geological type label.

[0058] In the geological feature classification model training process, the shield tunneling parameter data in the supervised learning data set is input into the trained geological feature extraction model for feature extraction, the output geological feature data is matched with the labeled geological type label, and then the geological feature classification model is trained. Embodiment

[0059] In a specific embodiment of the present application, a shield tunnel construction geological real-time prediction system based on a neural network is provided, which can realize the above-mentioned shield tunnel construction geological real-time prediction method based on a neural network, and comprises a data acquisition module, a data processing and storage module and a geological real-time prediction module, which realize data acquisition, data processing and storage and geological prediction functions, respectively.

[0060] The data acquisition module acquires shield tunneling parameter data and construction progress information in real time, inputs the required data information, and provides an information basis for the subsequent data processing and storage module. The acquired shield tunneling parameter data includes slurry pressure, pushing speed, pushing force, cutter head torque, cutter head speed and synchronous grouting pressure. The construction progress information includes tunneling time and tunneling ring number. In addition, the data acquisition module extracts geological information according to the geological survey report.

[0061] The data processing and storage module comprises three units of tunneling data preprocessing, geological type division and data storage. The module pre-processes and normalizes the obtained shield tunneling parameter data through the tunneling data preprocessing unit; then, through the geological type division unit, more than ten types of soil are clustered and divided according to the soil physical properties, providing data support for the subsequent training of the geological feature classification module; the processed data are stored as the data input of the geological feature extraction module. The data storage unit collects the data processed by the two units and systematically stores the data set.

[0062] A, the tunneling data preprocessing unit: the data obtained by the data acquisition module is cleaned, completed and removed. The main steps of this unit are abnormal value removal, data denoising, information compression and data normalization.

[0063] a) Outlier rejection: Outliers are defined as data points that deviate significantly from other data. These data points are generated by different factors such as sensors, shield tunneling machine itself and environmental factors, and should be rejected when data analysis is performed. This embodiment adopts the box plot method for outlier detection and rejection. Box plot is a statistical graphical description used to show the distribution of a set of data. It describes the distribution of data by plotting the maximum, minimum, median, first quartile (Q1) and third quartile (Q1). The interquartile range (IQR) in the box plot is defined as the difference between the third quartile (Q3) and the first quartile (Q1). In the process of outlier detection, any data point in the interval (-∞, Q1-1.5×IQR) and (Q3+1.5×IQR, +∞) is defined as an outlier and is rejected.

[0064] b) Data denoising: Data noise is interference data in the data set. Noise points deviate from normal data points, affecting the accuracy and calculation speed of the model. Noise points need to be filtered before model training. This embodiment uses wavelet transform for data denoising. Wavelet transform denoising removes noise in the signal through the decomposition and reconstruction process of wavelet transform. Wavelet transform denoising first selects a suitable wavelet basis function according to the characteristics of the signal and determines the number of wavelet decomposition layers. Then the signal is decomposed into multiple layers of wavelet, obtaining the wavelet decomposition coefficients at each scale. Then set the threshold, keep the low-frequency coefficients to keep the overall shape of the signal unchanged, and process the high-frequency coefficients with threshold. Finally, according to the processed wavelet coefficients, the signal is reconstructed to obtain the denoised signal estimate.

[0065] c) Information compression: The data acquisition module records the shield tunneling machine excavation parameters with time as the recording scale when collecting. The amount of original data obtained is huge, greatly increasing the calculation amount of the model. The excavation time is matched with the shield tunnel segment ring spacing, and the average value and standard deviation of the data within one ring spacing are used to replace the original data, realizing the compression of information.

[0066] d) Data normalization: Data normalization improves the convergence speed of model prediction and prevents numerical instability or gradient disappearance. This embodiment uses the extreme value normalization method to process the data, which linearly maps the data to the interval, and the calculation formula is: ,

[0067] x' is the data in the normalized data set; x is the data in the data set to be processed; x min is the minimum value of the data set; x max is the maximum value of the data set.

[0068] B, geological type division unit: cluster analysis is performed on soil types with similar physical properties to improve the prediction accuracy of the model and simplify the soil types while ensuring that it can provide guidance for shield tunneling parameter setting. The internal friction angle, cohesion and elastic modulus of each type of soil in the project are used as clustering analysis features for k-means++ unsupervised clustering analysis. First, an initial centroid is randomly selected from the sample set as the first cluster center c1. Then, for each sample point, the shortest distance (i.e., the distance to the nearest cluster center) D(x) between it and the selected cluster center is calculated, and the next cluster center c i is selected according to the probability of each sample point D(x). The farther the sample point, the greater the probability P(x) of being selected, so as to increase the dispersion of the cluster centers. Repeat the above steps until the required number of cluster centers (c1, c2,..., c k ) are selected. Finally, the selected initial cluster centers are used to perform the iteration process of the K-means clustering algorithm until convergence is reached or the stopping condition is met. According to the geological conditions in the shield construction interval, the soil is divided into 4 categories, and the division is shown in Table 1.

[0069] Table 1 Main soil division

[0070] Soil label Soil class Soil name 0 Muddy clay Muddy clay, muddy silty clay 1 Silty clay Silty clay, silty 2 Sand Silt, medium coarse sand, gravelly sand, moderately weathered argillaceous silt 3 Gravel Round gravel, moderately weathered psammite, strongly weathered psammite

[0071] C, the data obtained by the tunneling data preprocessing unit and the geological type division unit are linked according to the tunneling time and the number of tunneling rings to form a shield tunneling geological database, and are stored.

[0072] The geological real-time prediction module consists of three units: data set construction, geological feature extraction model and geological feature model classification. The data set construction unit divides the unsupervised learning data set and the supervised learning data set based on the shield tunneling geological database, and inputs data to the geological feature extraction unit and the geological feature classification unit, respectively. The geological feature extraction model and the geological feature classification model based on the integrated autoencoder and the residual CNN network realize the extraction and classification of geological features, respectively, and finally realize the prediction of geological types.

[0073] A, data set construction unit: in the shield tunneling geological database, due to the problem of insufficient comprehensive geological survey, the proportion of data with geological labels is small, and if the geological type prediction is directly performed, a large amount of data without geological labels will be wasted. Therefore, the data without geological labels are divided into an unsupervised learning data set for the construction of the geological feature extraction unit, and the data with geological labels are divided into a supervised learning data set for the construction of the geological feature classification unit. The proportions of the training set, the validation set and the test set of the two data sets are all 8:1:1. Each group of input feature data contains the average and standard deviation of the tunneling parameters of the previous 5 rings, and the tunneling parameter data at the prediction position.

[0074] B, geological feature extraction model: the model takes an unsupervised learning dataset as input, aiming to extract the geological information features in the input dataset and remove the redundant information in the data. The feature extraction model is constructed by an integrated autoencoder with an attention mechanism. The integrated autoencoder contains multiple integrated sub-autoencoders, and the core is to improve the robustness of feature extraction or reconstruction through the cooperative work of multiple sub-autoencoders. Multiple sub-autoencoders of different structures are independently trained, and the weights of the autoencoder are allocated through the attention mechanism Attention layer to fuse the output features and improve the effectiveness of feature extraction. The geological feature extraction model based on the integrated autoencoder is shown in FIG. 8, which contains three sub-autoencoders. Figure 1

[0075] Among them, the sub-autoencoder 1 is constructed based on the CNN network, which is used to extract the potential features of the intra-annulus and inter-annulus local space. The convolution kernel size of all Conv layers is 3, the sliding step and the padding layer number are 1, and the ReLU activation function is connected, and a BN layer is added between the convolution layer and the activation function. Three pooling layers are set in the encoder 1, the first layer uses the maximum pooling (MaxPool) to remove the redundant information in the data, and the last two layers use the average pooling (AvgPool) to retain all the information in the features. The window size and sliding step of the pooling layer are both 2. In the decoder 1, the inverse convolution ConvTranspose layer is used to perform up-sampling processing on the encoded features, the convolution kernel size is 3, the sliding step is 2, and the input padding layer number and the output padding layer are both 1. Finally, the output dimension of the encoded feature H1 is 1x16.

[0076] The sub-autoencoder 2 is constructed based on the LSTM network, which is used to extract the potential features in the time series. Four LSTM layers are set in the encoder 2, with neuron numbers of 128, 64, 32, and 16, and the LSTM layers are connected by Tanh activation functions. A random dropout layer is added at the end to prevent model overfitting, with a dropout coefficient of 0.2. Five LSTM layers are set in the decoder 2, with neuron numbers of 128, 64, 32, 16, and output dimension size, and the LSTM layers are connected by Tanh activation functions. Finally, the output dimension of the encoded feature H2 is 1x16.

[0077] ​The sub-autoencoder 3 is jointly constructed based on a CNN and an LSTM network, and is used to extract global latent features in the input data. The convolution kernel size of the Conv layer in the encoder 3 is 3, the sliding step and the padding layer number are 1, and the ReLU activation function is connected, and a BN layer is added between the convolution layer and the activation function. The first layer of the pooling layer uses maximum pooling (MaxPool), and the second layer uses average pooling (AvgPool). The window size and sliding step of the pooling layer are both 2. The number of neurons of the LSTM layer is 64 and 16, respectively, and the Tanh activation function is connected between each LSTM layer. A random inactivation Dropout layer with an inactivation coefficient of 0.2 is added at the end. In the decoder 3, an up-sampling Upsample layer is used for up-sampling processing, and the sizes are 32, 64 and 128, respectively, and the ReLU activation function is connected. Finally, the output dimension of the encoding feature H3 is 1x16.

[0078] Each autoencoder compresses the original input data into a latent space (hidden variable that compresses and quantizes the input data) through the construction of an encoder, and the decoder reconstructs the original input data according to the latent space. The loss function of the autoencoder is as follows:

[0079]

[0080] where n is the total sample size, is the i-th input sample, is the i-th predicted sample.

[0081] Through parallel training of the three autoencoders, the output integrated model encoding feature is obtained by splicing the sub-encoding features to form an encoding feature matrix , and a LeakyReLU activation function is introduced to calculate the attention score :

[0082]

[0083] where , is a learnable weight coefficient, is a learnable bias coefficient, and H i is an encoding feature, i=1, 2, 3. Then, the attention score calculation formula is updated by adding a global average pooling to obtain the upper and lower feature information:

[0084]

[0085] where c is the average pooling coefficient, and N is the total amount of model encoding.

[0086]

[0087] The output is converted into a probability distribution by Softmax linear regression, thereby generating normalized weights. Suppose the output of the original neural network is e1, e2,..., en n , then after Softmax regression processing, the weight a output is as follows:

[0088] ,

[0089] Finally, the normalized weights of each sub-encoder are used to weight and fuse the sub-encoding features, and the output dimension of the Attention encoding feature is 3x16 .

[0090] C. Geological feature classification model: The optimal parameters of the trained feature extraction model are then frozen, and a geological feature classification model based on a residual CNN neural network is established based on the geological feature extraction model. The feature extraction model encodes the supervised learning data set into geological features , and inputs them into the geological classification model to realize geological feature classification prediction. As shown in Figure 2 , the feature classification model is composed of residual convolution layers, convolution layers, fully connected layers, pooling layers, batch normalization layers and activation functions. The function of the fully connected layer is only to adjust the dimension of the input data. The role of the residual network is to alleviate the problem of gradient disappearance in the training process of the multi-layer network. The geological feature classification model contains two residual convolution layers, and the main structure inside is two convolution layers. The size of the convolution kernel of the convolution layer is 3, the sliding step and the number of padding layers are both 1. In all convolution layers except the residual convolution layer, the size of the convolution kernel of the second convolution layer is 4, the sliding step is 2, and the number of padding layers is 1. The size of the convolution kernel of the remaining convolution layers is 3, the sliding step and the number of padding layers are both 1, and all are connected using the ReLU activation function. Three pooling layers are set, the first two layers use MaxPool, and the last layer uses AvgPool. The window size and sliding step of the pooling layer are both 2. Finally, the classification of geological types is realized through the fully connected layer and the Softmax layer. The cross-entropy loss function is used to calculate the distance between the predicted probability distribution and the true probability distribution. The loss is calculated as follows:

[0091] .

[0092] where N and M are the total number of samples and the number of predicted categories, respectively; is the symbol function, which takes 1 if the true class of sample i is equal to c, and 0 otherwise; is the predicted probability that sample i belongs to class c.

[0093] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code thereon.

[0094] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0095] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0097] The above merely describes preferred embodiments of the present application, and the present application is not limited to the above. Any person skilled in the art, without departing from the technical scope of the present application, can make many possible changes and modifications to the technical solutions of the present application, or modify equivalent embodiments with equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, without departing from the technical scope of the present application, still falls within the protection scope of the technical solutions of the present application.

Claims

1.A neural network-based real-time prediction method for geological conditions in shield tunnel construction, characterized in that, The method comprises the following steps: obtaining and preprocessing shield tunneling parameter data, inputting the preprocessed shield tunneling parameter data into a pre-trained geological feature extraction model, and outputting geological feature data; then inputting the geological feature data into a pre-trained geological feature classification model, and outputting a geological category; the geological feature extraction model is a neural network model composed of an integrated autoencoder with an attention mechanism, and the integrated autoencoder comprises a plurality of parallel sub-autoencoders, each of which comprises an encoder and a decoder; the geological feature classification model is a residual CNN neural network model; the three parallel sub-autoencoders are respectively: a first sub-autoencoder based on a CNN network, used for extracting intra-annulus and inter-annulus spatial features; a second sub-autoencoder based on an LSTM network, used for extracting tunneling parameter time sequence features; a third sub-autoencoder based on a CNN-LSTM hybrid network, used for extracting global features. 2.The neural network-based real-time prediction method for shield tunneling geology according to claim 1, characterized in that, The shield tunneling parameter data comprises slurry pressure, pushing speed, pushing force, cutterhead torque, cutterhead rotating speed and synchronous grouting pressure; and the construction progress information comprises tunneling time and tunneling ring number. 3.The neural network-based real-time prediction method of shield tunneling geology according to claim 1, characterized in that, The preprocessing of the shield tunneling parameter data specifically comprises: performing abnormal value detection and elimination by using a box plot method; performing data denoising by using wavelet transform; then matching the tunneling time with the shield segment inter-annulus spacing, replacing the original data with the average value and the standard deviation of the data within one annulus spacing to realize information compression; and finally performing maximum and minimum normalization processing. 4.The neural network-based real-time prediction method for shield tunneling geology according to claim 1, characterized in that, The integrated autoencoder further comprises an attention fusion layer, and the three features are fused by the following steps: 1) concatenating the encoding features H1, H2 and H3 output by each sub-autoencoder into a feature matrix H; 2) calculating an attention score for each encoding feature by using a LeakyReLU activation function; 3) performing global average pooling operation on the attention score; 4) generating normalized weights by using a Softmax function, so that each weight value is in the interval [0, 1] and the sum is 1; 5) performing weighted summation according to the normalized weights of each encoding feature to output a fused feature. 5.The neural network-based real-time prediction method for shield tunneling geology according to claim 1, characterized in that, The pre-training method of the geological feature extraction model and the geological feature classification model is as follows: obtaining shield tunneling parameter data, construction progress information and geological survey report information; preprocessing the shield tunneling parameter data, clustering the geological types according to soil physical properties; constructing a shield tunneling parameter data set from the shield tunneling parameter data and the construction progress information; linking the preprocessed shield tunneling parameter data and the geological types according to the construction progress information to construct a shield tunneling geological data set; dividing the shield tunneling geological data set into an unsupervised learning data set and a supervised learning data set, training the geological feature extraction model using the unsupervised learning data set, and training each sub-autoencoder independently; then freezing the parameters of the geological feature extraction model, and training the geological feature classification model using the supervised learning data set. 6.The real-time prediction method of shield tunneling geology based on neural network according to claim 5, characterized in that, The clustering of the geological types according to the soil physical properties specifically comprises: taking the internal friction angle, cohesion and elastic modulus of each type of soil in the project as the characteristics of the clustering analysis based on the geological survey report information, performing K-means++ unsupervised clustering analysis, and dividing the soil into several geological types. 7.The neural network-based real-time prediction method of shield tunneling geology according to claim 5, characterized in that, The unsupervised learning data set is a shield tunneling parameter data set without a geological type label, and the supervised learning data set is a shield tunneling geological data set with a geological type label. 8.The neural network-based real-time prediction method for shield tunneling geology according to claim 7, characterized in that, In the training process of the geological feature classification model, shield tunneling parameter data in the supervised learning data set is input into the trained geological feature extraction model for feature extraction, the output geological feature data is matched with the labeled geological type label, and then the geological feature classification model is trained. 9.A neural network-based real-time geological prediction system for shield tunneling, characterized by, A method for implementing any one of claims 1-8, comprising: a data acquisition module for acquiring shield tunneling parameter data and preprocessing; a feature extraction module for inputting the preprocessed shield tunneling parameter data into a pre-trained geological feature extraction model to output geological feature data; a feature classification module for inputting the geological feature data output by the geological feature extraction model into a pre-trained geological feature classification model to output a geological category.

Images