Shield tunnel construction geology real-time prediction method and system based on neural network

Abstract

The invention discloses a shield tunnel construction geology real-time prediction method and system based on a neural network, and the method comprises the steps: obtaining and preprocessing tunneling parameter data of a shield tunneling machine, inputting the preprocessed tunneling parameter data of the shield tunneling machine into a pre-trained geologic feature extraction model, and outputting geologic feature data; inputting the geologic feature data into a pre-trained geologic feature classification model, and outputting a geologic category; the geological feature extraction model is a neural network model formed by an integrated auto-encoder added with an attention mechanism, and the integrated auto-encoder comprises a plurality of parallel sub-auto-encoders; the geologic feature classification model is a residual CNN neural network model. According to the method, real-time accurate geological prediction can be realized by fully utilizing the tunneling parameters of the whole construction section, the total sample size of data required by training is reduced, the accuracy and timeliness of model prediction are greatly improved, and the geological change condition can be more effectively judged.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel construction, and relates to a method and system for real-time prediction of tunnel construction geology, and in particular to a method and system for real-time prediction of shield tunnel construction geology based on a neural network. Background Art

[0002] With the rapid development of tunnel transportation, urban tunnel construction projects are gradually increasing. Shield tunnel construction, due to its low vibration, high operating efficiency, and minimal environmental impact, has gradually become the primary method for urban underground tunnel construction. However, during shield construction, due to the unclear geological conditions of the tunnel face, shield machine operators are unable to reasonably adjust the shield machine's excavation parameters. Improper operating parameters prevent the shield machine from adapting well to changes in geological conditions, resulting in abnormal shield machine excavation parameters and abnormal loading. Furthermore, the shield machine's cutter wear is severe, the excavation speed slows, and abnormal surface subsidence occurs, seriously affecting construction safety and progress. In order to improve construction efficiency and ensure the safety of construction and the surrounding environment, it is particularly important to accurately predict the geological information of the tunnel face.

[0003] After searching, the invention with Chinese patent number CN115659842A discloses a geological dynamic inversion method using a shield tunneling database, especially involving large-diameter shield machines, shield machines used for building underwater tunnels or construction in soft upper and hard lower strata. The excavation surface of large-diameter shield machines has complex strata, the geological survey of underwater tunnels is inaccurate, and the soil-rock interface of soft upper and hard lower strata is difficult to determine. The geological survey data and tunnel longitudinal section drawings of the project cannot accurately reflect the actual geological survey situation. By collecting and analyzing shield equipment data, slag information and information from several sensors, a data-driven stratum inversion is formed, and new electronic maps are continuously generated during shield tunneling to guide the implementation of shield construction. The invention with Chinese patent number CN108846521A proposes a method for predicting adverse geological types in shield construction based on Xgboost. The implementation steps are: preprocessing the PDV historical data of the shield machine; obtaining the key features of multiple tunneling parameter data after preprocessing; 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 types during shield construction.

[0004] However, the above-mentioned shield geological prediction related technologies do not fully consider the incompleteness of geological survey reports, which makes it impossible to fully utilize the tunneling parameter data, and the accuracy and timeliness of geological prediction are low.

[0005] Existing technologies for geological prediction during shield tunneling only utilize tunneling parameter data from sections with geological drilling, failing to fully consider the incompleteness of geological survey reports. This prevents the full utilization of tunneling parameter data from sections without geological drilling. When localized geological changes occur between adjacent geological drilling intervals, existing technologies are unable to accurately predict the geological type based on the changes in tunneling parameters, resulting in low geological prediction accuracy and timeliness. Furthermore, the large amount of tunneling parameter data cannot be fully utilized, requiring a larger total sample size for network training.

[0006] After searching, the invention with Chinese patent number CN113431635A discloses a semi-supervised method and system for predicting the geological type of the shield tunnel face, including the following steps: screening out machine parameters related to geological conditions from machine data; preprocessing the screened machine parameters; constructing an unlabeled data set and a labeled data set based on the preprocessed machine parameters; establishing a semi-supervised framework for predicting the geological conditions of the tunnel face to obtain a geological feature extractor and a feature classifier; using a constrained DenseNet autoencoder network and an unlabeled data set to train the geological feature extractor; using a deep neural network and a labeled data set to train the geological feature classifier, and finally realizing the prediction of the geological type ahead of the face.

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

[0008] Therefore, a method is urgently needed to fully utilize the tunneling parameter data of the entire construction section for geological prediction and identification, so as to reduce the total sample size of training data, improve the accuracy and timeliness of geological prediction during shield construction, and provide timely guidance for the active parameter setting of the shield machine. Summary of the Invention

[0009] To address these issues, the present invention proposes a neural network-based real-time geological prediction method and system for shield tunneling. This patented method leverages the tunneling parameters of shield machines to extract geological characteristics for geological prediction, thereby guiding the setting of shield machine operating parameters, improving construction progress, and reducing construction risks.

[0010] The technical solution adopted in the present invention is as follows:

[0011] A neural network-based real-time prediction method for shield tunnel construction geology includes the following steps:

[0012] Obtaining shield machine excavation parameter data and preprocessing it, inputting the preprocessed shield machine excavation parameter data into a pretrained geological feature extraction model to output geological feature data; then inputting the geological feature data into a pretrained geological feature classification model to output a geological category;

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

[0014] Furthermore, the shield machine excavation parameter data includes mud water pressure, propulsion speed, propulsion force, cutter head torque, cutter head rotation speed and synchronous grouting pressure; the construction progress information includes excavation time and number of excavation rings.

[0015] Furthermore, the preprocessing of the shield machine excavation parameter data specifically includes: using the box plot method to detect and eliminate outliers; using wavelet transform to perform data denoising; then matching the excavation time with the shield machine segment ring spacing, replacing the original data with the average value and standard deviation of the data within a ring spacing to achieve information compression; and finally performing maximum value normalization processing.

[0016] Furthermore, the three parallel sub-autoencoders are:

[0017] The first sub-autoencoder based on the CNN network is used to extract intra-ring and inter-ring spatial features;

[0018] The second sub-autoencoder based on the LSTM network is used to extract the time series features of tunneling parameters;

[0019] The third sub-autoencoder based on the CNN-LSTM hybrid network is used to extract global features.

[0020] Furthermore, the integrated autoencoder also includes an attention fusion layer, which fuses the three features through the following steps:

[0021] 1) Concatenate the encoded features H1, H2, and H3 output by each sub-autoencoder into a feature matrix H;

[0022] 2) For each encoded feature, the attention score is calculated through the LeakyReLU activation function;

[0023] 3) Perform global average pooling on the attention scores;

[0024] 4) The pooled attention scores are passed through the Softmax function to generate normalized weights so that each weight value is in the range [0, 1] and the sum is 1;

[0025] 5) Perform weighted summation according to the normalized weights of each encoding feature and output the fused feature.

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

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

[0028] Preprocess the shield machine excavation parameter data and cluster the geological types according to the soil physical properties; construct a shield machine excavation parameter dataset by combining the shield machine excavation parameter data and construction progress information; and link the preprocessed shield machine excavation parameter data and geological types based on the construction progress information to construct a shield machine excavation geological dataset.

[0029] The shield tunneling geological dataset is divided into an unsupervised learning dataset and a supervised learning dataset. The geological feature extraction model is trained using the unsupervised learning dataset, and each sub-autoencoder is trained independently. Then, the parameters of the geological feature extraction model are frozen, and the geological feature classification model is trained using the supervised learning dataset.

[0030] Furthermore, clustering geological types according to 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 features of cluster analysis, performing K-means++ unsupervised cluster analysis, and dividing the soil into several geological types.

[0031] Furthermore, the unsupervised learning dataset is a shield tunneling parameter dataset without geological type labels, and the supervised learning dataset is a shield tunneling geological dataset with geological type labels.

[0032] Furthermore, during the training process of the geological feature classification model, the shield machine excavation 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 marked geological type label, and then the geological feature classification model is trained.

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

[0034] Data acquisition module: used to obtain shield machine excavation parameter data and perform preprocessing;

[0035] Feature extraction module: used to input the pre-processed shield machine excavation parameter data into the pre-trained geological feature extraction model and output geological feature data;

[0036] Feature classification module: used to input the geological feature data output by the geological feature extraction model into the pre-trained geological feature classification model and output the geological category. Compared with the existing technology, the beneficial effects of this invention are:

[0037] Current geological prediction technology for shield tunneling construction relies solely on a single feature extraction model to simultaneously extract features from multi-dimensional data. This results in low feature extraction effectiveness and insufficient prediction model accuracy. Compared to existing technologies, the real-time geological prediction method proposed in this paper can extract multi-dimensional features of input data, such as time and space, based on feature extraction models with different functions. This significantly improves the accuracy and timeliness of model predictions and enables more effective identification of geological changes. This provides guidance for shield tunneling operators in setting active shield parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of a geological feature extraction model based on an integrated autoencoder with an attention mechanism in an embodiment of the present invention.

[0039] Figure 2 Schematic diagram of a geological feature classification model based on a convolutional residual neural network in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The technical solution of the present invention will be further explained in detail below with reference to the accompanying drawings and specific embodiments.

[0041] A neural network-based real-time prediction method for shield tunnel construction geology includes the following steps:

[0042] Shield machine excavation parameter data is obtained and preprocessed, and the preprocessed shield machine excavation parameter data is input into a pre-trained geological feature extraction model to output geological feature data; then the geological feature data is input into a pre-trained geological feature classification model to output a geological category.

[0043] The shield machine excavation parameter data includes slurry pressure, propulsion speed, propulsion force, cutterhead torque, cutterhead speed, and synchronous grouting pressure; the construction progress information includes excavation time and number of excavation rings. The preprocessing of the shield machine excavation parameter data specifically includes: using a box plot method to detect and eliminate outliers; using wavelet transform to perform data denoising; then matching the excavation time with the shield machine segment ring spacing, replacing the original data with the mean and standard deviation of the data within a ring spacing to achieve information compression; and finally performing maximum value normalization.

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

[0045] The three parallel sub-autoencoders are: the first sub-autoencoder based on the CNN network, used to extract the spatial features within and between rings; the second sub-autoencoder based on the LSTM network, used to extract the temporal features of the tunneling parameters; and the third sub-autoencoder based on the CNN-LSTM hybrid network, used to extract the global features.

[0046] The integrated autoencoder also includes an attention fusion layer, which fuses the three-way features through the following steps:

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

[0048] 2) For each encoded feature, the attention score is calculated through the LeakyReLU activation function;

[0049] 3) Perform global average pooling on the attention scores;

[0050] 4) The pooled attention scores are passed through the Softmax function to generate normalized weights so that each weight value is in the range [0, 1] and the sum is 1;

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

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

[0053] Obtain shield machine excavation parameter data, construction progress information and geological survey report information;

[0054] Preprocess the shield machine excavation parameter data and cluster the geological types according to the soil physical properties; construct a shield machine excavation parameter dataset by combining the shield machine excavation parameter data and construction progress information; and link the preprocessed shield machine excavation parameter data and geological types based on the construction progress information to construct a shield machine excavation geological dataset.

[0055] The shield tunneling geological dataset is divided into an unsupervised learning dataset and a supervised learning dataset. The geological feature extraction model is trained using the unsupervised learning dataset, and each sub-autoencoder is trained independently. Then, the parameters of the geological feature extraction model are frozen, and the geological feature classification model is trained using the supervised learning dataset.

[0056] The clustering of geological types according to soil physical properties specifically includes: based on the geological survey report information, the internal friction angle, cohesion and elastic modulus of each type of soil in the project are used as cluster analysis features, and K-means++ unsupervised cluster analysis is performed to divide the soil into several geological types.

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

[0058] During the training process of the geological feature classification model, the shield machine excavation 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 marked geological type label, and then the geological feature classification model is trained. Example

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

[0060] The data acquisition module acquires shield machine excavation parameter data and construction progress information in real time, inputting the required data and information to provide an information foundation for subsequent data processing and storage modules. Collected shield machine excavation parameters include slurry pressure, propulsion speed, propulsion force, cutterhead torque, cutterhead speed, and synchronous grouting pressure. Construction progress information includes excavation time and number of excavation cycles. Furthermore, the data acquisition module extracts geological information based on geological survey reports.

[0061] The data processing and storage module comprises three units: tunneling data preprocessing, geological classification, and data storage. The tunneling data preprocessing unit preprocesses and normalizes the shield machine tunneling parameter data obtained. The geological classification unit then clusters more than ten soil types based on their physical properties, providing data support for the subsequent training of the geological feature classification module. The processed data is stored as input for the geological feature extraction module. The data storage unit collects the data processed by these two units and systematically stores the dataset.

[0062] A. Tunneling Data Preprocessing Unit: This unit cleans, completes, and eliminates data acquired by the data acquisition module. The main steps of this unit are outlier removal, data denoising, information compression, and data normalization.

[0063] a) Outlier Removal: Outliers are defined as data points whose values ​​significantly deviate from the rest of the data. These data points are generated by various factors, including sensors, the shield machine itself, and environmental factors. These data points should be removed during data analysis. This embodiment uses a box plot method for outlier detection and removal. A box plot is a statistical graphical representation used to display the distribution of a set of data. It depicts the data distribution by plotting the maximum, minimum, median, first quartile (Q1), and third quartile (Q2). The interquartile range (IQR) in a box plot is defined as the difference between the third quartile (Q3) and the first quartile (Q1). During outlier detection, any data point within the interval (-∞, Q1 - 1.5 × IQR) and (Q3 + 1.5 × IQR, +∞) is defined as an outlier and removed.

[0064] b) Data denoising: Data noise is the interfering data in the data set. Noise points deviate from normal data points, thus affecting the accuracy and calculation speed of the model. Noise points need to be filtered out before model training. This embodiment uses wavelet transform for data denoising. Wavelet transform denoising removes noise from 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 layers of wavelet decomposition. The signal is then subjected to multi-layer wavelet decomposition to obtain wavelet decomposition coefficients at each scale. Then, a threshold is set, the low-frequency coefficients are retained to keep the overall shape of the signal unchanged, and the high-frequency coefficients are thresholded. Finally, the signal is reconstructed based on the processed wavelet coefficients to obtain the denoised signal estimate.

[0065] c) Information Compression: The data acquisition module acquires shield machine excavation parameters using time as the record scale. This produces a large amount of raw data, significantly increasing the model's computational complexity. By matching excavation time with the spacing between shield machine segments, the raw data is replaced with the mean and standard deviation of the data within a ring spacing, achieving information compression.

[0066] d) Data normalization: Data normalization is used to improve the convergence speed of model prediction and prevent numerical instability or gradient vanishing problems. This embodiment uses maximum value normalization to linearly map the data to an interval. The calculation formula is: ,

[0067] x' is the data in the normalized data set; x is the data in the excavation parameter data set to be processed; x min is the minimum value of the tunneling parameter data set; x max It is the maximum value of this set of tunneling parameter data sets.

[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 the setting of shield tunneling parameters. The internal friction angle, cohesion and elastic modulus of each type of soil in the project are used as the characteristics of cluster analysis, and k-means++ unsupervised cluster analysis is performed. 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 between it and the selected cluster center (that is, the distance to the nearest cluster center) D(x) is calculated and the next cluster center c is selected according to the probability based on D(x) of each sample point. i The farther the sample point is, the greater the probability P(x) of being selected is, 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 Finally, the K-means clustering algorithm is iterated using the selected initial cluster centers until convergence is achieved or a stopping condition is met. Based on the geological conditions within the shield construction area, the soil is divided into four major categories, as shown in Table 1.

[0069] Table 1 Main soil classification

[0070] Soil Label Soil type Soil name 0 silty clay silty clay, silty clay 1 silty clay Silty clay, silty clay 2 sand and gravel silt, medium-coarse sand, gravel sand, medium-weathered muddy silt 3 gravel Gravel, moderately weathered conglomerate, strongly weathered conglomerate

[0071] C. Link the data obtained from the tunneling data preprocessing unit and the geological type classification unit based on the tunneling time and the number of tunneling rings to form a shield tunneling geological database and store it.

[0072] The real-time geological prediction module consists of three units: dataset construction, geological feature extraction model, and geological feature model classification. The dataset construction unit divides the unsupervised learning dataset and the supervised learning dataset based on the shield tunneling geological database, which serve as input for the geological feature extraction unit and the geological feature classification unit, respectively. The geological feature extraction model and the geological feature classification model respectively extract and classify geological features using an integrated autoencoder and a residual CNN network, ultimately achieving geological type prediction.

[0073] A. Dataset Construction Unit: Due to the lack of comprehensive geological surveys in shield tunneling geological databases, data with geological labels accounts for a relatively small proportion. Directly predicting geological types would waste a significant amount of unlabeled data. Therefore, the unlabeled data is segmented into an unsupervised learning dataset for geological feature extraction, while the labeled data is segmented into a supervised learning dataset for geological feature classification. The training, validation, and test sets for both datasets have a ratio of 8:1:1. Each set of input feature data contains the mean and standard deviation of the tunneling parameters for the first five rings, as well as the tunneling parameter data for the predicted location.

[0074] B. Geological feature extraction model: This model takes an unsupervised learning dataset as input, and aims to extract geological information features from the input dataset and remove redundant information from 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 its core is to improve the robustness of feature extraction or reconstruction through the collaborative work of multiple sub-autoencoders. Multiple sub-autoencoders with different structures are trained independently, and the autoencoder weights are assigned through the attention mechanism Attention layer, and the output features are fused to improve the effectiveness of feature extraction. Geological feature extraction model based on integrated autoencoder Figure 1 As shown in Figure 2, there are three sub-autoencoders in total, each of which consists of its own encoder and decoder.

[0075] Sub-autoencoder 1 is built on a CNN network and is used to extract latent features in the local space within and between rings. All Conv layers have a convolution kernel size of 3, a sliding stride and a padding layer of 1. Reluctant Luminance (ReLU) activation functions are used for connection, and a batch normalization (BN) layer is added between the convolution layers and the activation functions. Encoder 1 has three pooling layers. The first layer uses max pooling (MaxPool) to remove redundant information from the data, while the next two layers use average pooling (AvgPool) to preserve all information in the features. The window size and sliding stride of the pooling layers are both 2. Decoder 1 uses a reverse convolution (ConvTranspose) layer to upsample the encoded features. The convolution kernel size is 3, the sliding stride is 2, and the input and output padding layers are both 1. The final output is the encoded feature H1 with a dimension of 1×16.

[0076] Sub-autoencoder 2 is built based on an LSTM network and is used to extract latent features from time series. Encoder 2 has four LSTM layers with 128, 64, 32, and 16 neurons, respectively. Tanh activation is used to connect the LSTM layers. A dropout layer with a dropout coefficient of 0.2 is added at the end to prevent overfitting. Decoder 2 has five LSTM layers with 128, 64, 32, and 16 neurons, respectively. Tanh activation is used to connect the LSTM layers. The final output is the encoded feature H2 with a dimension of 1×16.

[0077] Sub-autoencoder 3 is constructed based on a CNN and LSTM network, extracting global latent features from the input data. In encoder 3, the convolution kernel size of the Conv layer is 3, the sliding stride and the number of padding layers are 1, and both layers use the Reluctant Unit (ReLU) activation function. A batch normalization (BN) layer is added between the convolution layers and the activation function. The first pooling layer uses max pooling (MaxPool), and the second layer uses average pooling (AvgPool). The window size and sliding stride of the pooling layer are both 2. The LSTM layers have 64 and 16 neurons, respectively, and are connected using the Tanh activation function. A dropout layer with a dropout coefficient of 0.2 is added at the end. In decoder 3, upsampling is performed using upsample layers with kernel sizes of 32, 64, and 128, respectively, connected using the Reluctant Unit (ReLU) activation function. The final output is the encoded feature H3 with a dimension of 1×16.

[0078] Each autoencoder compresses the original input data into a latent space (compressing and quantizing the latent variables of the input data) by constructing an encoder, and the decoder then reconstructs the original input data based on the latent space. The loss function of the autoencoder is shown below:

[0079]

[0080] Where n is the total sample size, is the i-th input sample, is the i-th prediction sample.

[0081] Through the parallel training of three autoencoders, the integrated model encoding features are output , concatenate the sub-encoding features to form the encoding feature matrix , introduce the LeakyReLU activation function to calculate the attention score :

[0082]

[0083] in 、 is the learnable weight coefficient, is the learnable bias coefficient, H i is the encoding feature, i = 1, 2, 3. Then the attention score calculation formula is updated by adding 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 through Softmax linear regression to generate normalized weights. Assume that the output of the original neural network is e1, e2, ..., e n , then after Softmax regression processing, the weight α output is as follows:

[0088] ,

[0089] Finally, the sub-encoding features are weighted and fused through the normalized weights of each sub-autoencoder, and the output dimension is 3×16 Attention encoding features. .

[0090] C. Geological feature classification model: The optimal parameters of the feature extraction model obtained through training are then frozen, and a geological feature classification model based on the residual CNN neural network is established based on the geological feature extraction model. The feature extraction model encodes the supervised learning dataset into geological features. , and input it into the geological classification model to achieve geological feature classification prediction. Figure 2 As shown in the figure, this feature classification model consists of a residual convolutional layer, a convolutional layer, a fully connected layer, a pooling layer, a batch normalization layer, and an activation function. The fully connected layer only adjusts the dimensionality of the input data. The residual network is used to alleviate the vanishing gradient problem during multi-layer network training. This geological feature classification model contains two residual convolutional layers, the main internal structure of which is two convolutional layers. The convolution kernel size of each convolutional layer is 3, and the sliding stride and number of padding layers are both 1. In all convolutional layers except the residual convolution layer, the second convolution layer has a convolution kernel size of 4, a sliding stride of 2, and a padding layer of 1. The remaining convolution layers all have a convolution kernel size of 3, a sliding stride and a padding layer of 1, and are all connected using the ReLU activation function. Three pooling layers are set, with the first two using MaxPool and the last using AvgPool. The window size and sliding stride of the pooling layer are both 2. Finally, geological type classification is achieved through a fully connected layer and a Softmax layer. The cross entropy loss function is used to calculate the distance between the predicted probability distribution and the true probability distribution. The cross entropy loss function calculates the loss as follows:

[0091] .

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

[0093] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts 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, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0095] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0097] The above description is only a preferred embodiment of the present invention. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can use the above disclosed methods and technical contents to make many possible changes and modifications to the technical solution of the present invention without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment with equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A neural network-based real-time prediction method for shield tunnel construction geology, characterized in that: The following steps are involved: Obtaining shield machine excavation parameter data and preprocessing it, inputting the preprocessed shield machine excavation parameter data into a pretrained geological feature extraction model to output geological feature data; then inputting the geological feature data into a pretrained geological feature classification model to output a geological category; The geological feature extraction model is a neural network model composed of an integrated autoencoder with an attention mechanism, wherein the integrated autoencoder includes multiple parallel sub-autoencoders, each of which includes an encoder and a decoder; the geological feature classification model is a residual CNN neural network model.

2. The neural network-based real-time prediction method for shield tunnel construction geology according to claim 1 is characterized in that: The shield machine excavation parameter data includes mud water pressure, propulsion speed, propulsion force, cutter head torque, cutter head rotation speed and synchronous grouting pressure; the construction progress information includes excavation time and number of excavation rings.

3. The neural network-based real-time prediction method for shield tunnel construction geology according to claim 1 is characterized in that: The preprocessing of shield machine excavation parameter data specifically includes: using a box plot method to detect and eliminate outliers; using wavelet transform to perform data denoising; then matching the excavation time with the shield machine segment ring spacing, replacing the original data with the average value and standard deviation of the data within a ring spacing to achieve information compression; and finally performing maximum value normalization processing.

4. The neural network-based real-time prediction method for shield tunnel construction geology according to claim 1, characterized in that: The three parallel sub-autoencoders are: The first sub-autoencoder based on the CNN network is used to extract intra-ring and inter-ring spatial features; The second sub-autoencoder based on the LSTM network is used to extract the time series features of tunneling parameters; The third sub-autoencoder based on the CNN-LSTM hybrid network is used to extract global features.

5. The neural network-based real-time prediction method for shield tunnel construction geology according to claim 1, characterized in that: The integrated autoencoder also includes an attention fusion layer, which fuses the three-way features through the following steps: 1) Concatenate the encoded features H1, H2, and H3 output by each sub-autoencoder into a feature matrix H; 2) For each encoded feature, the attention score is calculated through the LeakyReLU activation function; 3) Perform global average pooling on the attention scores; 4) The pooled attention scores are passed through the Softmax function to generate normalized weights so that each weight value is in the range [0, 1] and the sum is 1; 5) Perform weighted summation according to the normalized weights of each encoding feature and output the fused feature.

6. The neural network-based real-time prediction method for shield tunnel construction 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: Obtain shield machine excavation parameter data, construction progress information and geological survey report information; Preprocess the shield machine excavation parameter data and cluster the geological types according to the soil physical properties; construct a shield machine excavation parameter dataset by combining the shield machine excavation parameter data and construction progress information; and link the preprocessed shield machine excavation parameter data and geological types based on the construction progress information to construct a shield machine excavation geological dataset. The shield tunneling geological dataset is divided into an unsupervised learning dataset and a supervised learning dataset. The geological feature extraction model is trained using the unsupervised learning dataset, and each sub-autoencoder is trained independently. Then, the parameters of the geological feature extraction model are frozen, and the geological feature classification model is trained using the supervised learning dataset.

7. The neural network-based real-time prediction method for shield tunnel construction geology according to claim 6, characterized in that: The clustering of geological types according to soil physical properties specifically includes: based on the geological survey report information, the internal friction angle, cohesion and elastic modulus of each type of soil in the project are used as cluster analysis features, and K-means++ unsupervised cluster analysis is performed to divide the soil into several geological types.

8. The neural network-based real-time prediction method for shield tunnel construction geology according to claim 6, characterized in that: The unsupervised learning dataset is a shield tunneling parameter dataset without geological type labels, and the supervised learning dataset is a shield tunneling geological dataset with geological type labels.

9. The neural network-based real-time prediction method for shield tunnel construction geology according to claim 8, characterized in that: During the training process of the geological feature classification model, the shield machine excavation 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 marked geological type label, and then the geological feature classification model is trained.

10. A real-time geological prediction system for shield tunnel construction based on neural network, characterized in that: The method for implementing any one of claims 1 to 9 comprises: Data acquisition module: used to obtain shield machine excavation parameter data and perform preprocessing; Feature extraction module: used to input the pre-processed shield machine excavation parameter data into the pre-trained geological feature extraction model and output geological feature data; Feature classification module: used to input the geological feature data output by the geological feature extraction model into the pre-trained geological feature classification model and output the geological category.

Images