Adverse geological type identification method and system based on joint deep learning network

Abstract

The application provides a joint deep learning network-based adverse geological type identification method and system, and relates to the technical field of geological disaster prediction.The method comprises the following steps: obtaining training information, including geological data of an excavation section, geophysical exploration inversion imaging information and drilling information; inputting the training information into a pre-established joint deep learning network model, sequentially performing data preprocessing, feature extraction and feature fusion to obtain joint feature representation, and obtaining the identification type and the form description of the adverse geological body based on the joint feature representation; wherein the joint deep learning network model comprises a DNN model, a CNN model and an RNN model, and the DNN model, the CNN model and the RNN model are used to extract the features of digital data, image data and text description data in the training information respectively.The application can quickly and accurately identify the type and form of the adverse geological body in front of a tunnel, thereby reducing the pressure of manual prediction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of geological disaster prediction, and particularly relates to a method and system for identifying an unfavorable geological type based on a joint deep learning network. BACKGROUND

[0002] In the field of tunnel engineering, the existence of unfavorable geological bodies poses many challenges and hidden dangers to the construction process. If the unfavorable geological bodies cannot be accurately predicted and targeted measures are not taken, it will cause unpredictable harm to the safety of tunnel construction personnel, greatly prolong the construction period, and result in more cost investment. For example, in the tunnel engineering in the broken zone stratum, collapse and other problems often occur due to the broken rock mass. If the broken zone is water-rich, the excavation process will also face the risk of water and mud inrush. Therefore, it is very important and necessary to predict the unfavorable geology in front of the tunnel, classify the unfavorable geology, depict the unfavorable geology form, and take targeted measures in tunnel construction.

[0003] Due to the development of detection means, the current position prediction technology for unfavorable geological bodies is very mature. However, the technology for judging the type and characteristics of unfavorable geological bodies and guiding excavation and risk disposal according to the type and characteristics is still immature and still highly dependent on expert team judgment.

[0004] The expert team has high credibility in type identification and feature description of unfavorable geological bodies, but the overall identification still relies on manual work, and the results of type judgment and feature description vary from person to person without a unified standard for identification. Moreover, due to the limitations of brain computing, the expert team cannot fuse multiple information when identifying the type and characteristics of unfavorable geological bodies, but only rely on memory and experience to associate several information for identification.

[0005] In addition, there are few expert teams that can identify the type and characteristics of unfavorable geological bodies, and when there is no expert team on site, construction personnel cannot accurately identify the type and characteristics of unfavorable geological bodies in front of the tunnel. SUMMARY

[0006] To overcome the above-mentioned deficiencies of the prior art, the present application provides a method and system for identifying an unfavorable geological type based on a joint deep learning network, which can quickly and accurately classify the type and depict the form of unfavorable geological bodies in front of the tunnel, making the geological prediction intelligent and greatly reducing the pressure of manual prediction.

[0007] To achieve the above-mentioned purpose, one or more embodiments of the present application provide the following technical solutions:

[0008] The present application provides a method for identifying an unfavorable geological type based on a joint deep learning network.

[0009] The adverse geological type identification method based on a joint deep learning network comprises the following steps:

[0010] Training information is obtained, including geological data of a tunneling section, geophysical inversion imaging information and drilling information;

[0011] The training information is input into a pre-established joint deep learning network model, data preprocessing, feature extraction and feature fusion are sequentially performed, joint feature representation is obtained, and the identification type and morphological description of the adverse geological body are obtained based on the joint feature representation;

[0012] The joint deep learning network model comprises a DNN model, a CNN model and an RNN model arranged side by side, and the features of digital data, image data and text description data in the training information are extracted by the DNN model, the CNN model and the RNN model respectively when the features are extracted.

[0013] The second aspect of the present application provides an adverse geological type identification system based on a joint deep learning network.

[0014] The adverse geological type identification system based on a joint deep learning network comprises:

[0015] The data acquisition module is configured to obtain training information, including geological data of a tunneling section, geophysical inversion imaging information and drilling information;

[0016] The model prediction module is configured to input the training information into a pre-established joint deep learning network model, sequentially perform data preprocessing, feature extraction and feature fusion, obtain joint feature representation, and obtain the identification type and morphological description of the adverse geological body based on the joint feature representation.

[0017] The joint deep learning network model comprises a DNN model, a CNN model and an RNN model arranged side by side, and the features of digital data, image data and text description data in the training information are extracted by the DNN model, the CNN model and the RNN model respectively when the features are extracted.

[0018] The above one or more technical solutions have the following beneficial effects:

[0019] 1. The present application proposes an adverse geological type identification method and system based on a joint deep learning network, which processes multi-source information through a joint deep learning network to more quickly and accurately identify the type and morphology of adverse geological bodies in front of a tunnel.

[0020] 2、The joint deep learning network is adopted, multiple deep learning networks are combined, a specific deep learning network is selected according to a specific data type, and when feature extraction is carried out, the DNN model, the CNN model and the RNN model are used to extract the features of the digital data, the image data and the text description data in the training information, so that the accuracy of multi-source information feature extraction is greatly improved.

[0021] 3、The joint deep learning network is adopted to depict the morphological characteristics of the adverse geological body, the natural language generation model is accessed to generate text description, and the generative adversarial network is accessed to draw a concept map, so that the adverse geological conditions in front of the tunnel can be more intuitively and clearly understood, and more accurate treatment measures can be made.

[0022] 4、When the model is trained, the loss function used for training the adverse geological body type discrimination and the adverse geological body shape description is different, the loss function for training the adverse geological body type discrimination includes a cross-entropy function, wherein a physical constraint can be added, and the loss function for training the adverse geological body shape description includes a mean square error function; the loss function for adverse geological body type discrimination and the loss function for adverse geological body shape description are weighted, a joint loss function of the joint deep learning network model is formed, and thus the overall training of the adverse geological body type discrimination and the adverse geological body shape description is realized.

[0023] The advantages of the additional aspects of the application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0024] The accompanying drawings, which form a part of the specification, are included to provide a further understanding of the application and are incorporated herein by reference. The illustrations are shown for the purpose of enabling those skilled in the art to implement the application and are not intended to limit the scope of the application.

[0025] Figure 1 The method flowchart of the first embodiment. DETAILED DESCRIPTION

[0026] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs.

[0027] It should be noted that the terms used herein are for the purpose of describing specific embodiments and are not intended to limit the scope of the exemplary embodiments according to the application.

[0028] In the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other.

[0029] Embodiment one

[0030] Regarding the current identification of the types and characteristics of unfavorable geological bodies mentioned in the background art, and the guidance of excavation and risk management based on these characteristics, there is a high reliance on the judgment of expert teams. However, these expert teams often rely only on memory and experience to link several pieces of information for identification, failing to effectively integrate multiple pieces of information, and are highly subjective. To address this, this embodiment discloses a method for identifying unfavorable geological types based on a joint deep learning network. By utilizing the joint deep learning network to process multi-source information in the forecast segment, it can achieve rapid and accurate forecasting of unfavorable geological bodies.

[0031] like Figure 1 The figure shows a framework diagram of a joint deep learning network. The method for identifying adverse geological types based on a joint deep learning network proposed in this embodiment may specifically include the following steps:

[0032] (1) Establish a database of adverse geological bodies.

[0033] (2) The geological data, geophysical inversion imaging information and drilling information of the excavated sections are screened and summarized and organized into training information.

[0034] (3) Establish a joint deep learning network input layer to preprocess the training information.

[0035] (4) Establish a joint deep learning network extraction layer and use different deep learning network models to extract features of different types of information in geological data, geophysical inversion imaging information, and drilling information.

[0036] Among them, the joint deep learning network model includes a DNN model, a CNN model and a RNN model set in parallel. When performing feature extraction, the DNN model, the CNN model and the RNN model are used to extract the features of digital data, image data and text description data in the training information respectively.

[0037] (5) A joint deep learning network fusion layer is established, and the attention mechanism is used to fuse the features of the extracted geological information, geophysical inversion imaging information, and drilling information to obtain a joint feature representation.

[0038] (6) Establish a joint deep learning network output layer to perform type discrimination and morphological characterization of adverse geological bodies based on the obtained joint feature representation.

[0039] (7) The above model is trained in a way that minimizes the loss function.

[0040] (8) After training is completed, the geological information, geophysical inversion information, and drilling information of the forecast period are input into the joint deep learning network to obtain the prediction results;

[0041] The prediction results output by the joint deep learning network are compared with the results of the adverse geological body database to complete the identification of the adverse geological type and the morphological description of the prediction section.

[0042] The prediction results of the joint deep learning network are compared with the results of the adverse geological body database to verify the reliability of the prediction results of the joint deep learning network.

[0043] Further, in step (1), the database is established based on existing adverse geological body information, and the database information covers the type division of adverse geology and the feature description of adverse geology.

[0044] The types of adverse geology include but are not limited to fault fracture zone, karst, weak rock stratum, water-rich fault and aquifer, coal seam, permafrost, etc.

[0045] The feature description of adverse geology includes the description of adverse geology including but not limited to adverse geology type, development condition (influence range, scale size) and possible geological disasters.

[0046] In step (2), the geological data can specifically include but is not limited to main lithological composition description, underground water distribution description, main large fault distribution description, geological profile map and other related geological maps; the geophysical inversion imaging information can specifically include but is not limited to seismic inversion imaging information, resistivity method inversion imaging information, induced polarization method inversion imaging information, transient electromagnetic method inversion imaging information, and geological radar imaging information; the drilling information can specifically include but is not limited to core picture, core columnar diagram, drilling information, etc.

[0047] The geological data, geophysical inversion imaging information and drilling information of the excavated section are screened, and the specific screening means can be:

[0048] (1) For geological data, first, representative lithological parameters such as rock type, strength, weathering degree, etc. can be screened out by analyzing the lithological data of the excavated section. Data with special geological structures (such as faults and folds) are highlighted. Secondly, for structure information related to adverse geology, such as fault fracture zone, landslide body, karst cave, etc., detailed descriptions of these features are screened out and distinguished from normal geological sections. Thirdly, by comparing the geological information of different periods and different positions, the parts with significant differences from normal strata (such as porosity and rock stratum inclination) are screened out as the focus of study. In addition, by statistically analyzing a large amount of geological data, factors that have a greater impact on tunnel construction safety and stability are screened out, and repeated or irrelevant data are removed.

[0049] (2) For geophysical inversion imaging information, first, use inversion imaging technology (such as resistivity imaging, seismic wave reflection, etc.) to identify abnormal bodies in the stratum, filter out inversion results related to faults, fracture zones, karst caves and other unfavorable geology, and exclude noise and abnormal interference. Second, for the results of multiple geophysical methods (such as electrical method, seismic exploration, radar), high-quality signals in a specific frequency band can be selected through spectrum analysis, band filtering and other means, and irrelevant or repetitive information can be removed. Third, based on the physical property differences (such as resistivity, reflection coefficient, etc.) of different geological bodies in the imaging results, areas that meet the characteristics of unfavorable geology types can be selected, and areas consistent with normal strata can be filtered out.

[0050] (3) For drilling information: first, analyze the core samples of the drilling section, select samples related to unfavorable geology (such as fracture development, fragmentation, water-rich layer), and record key indicators in the core, such as rock mass integrity and water content. Second, combined with various geological parameters obtained during drilling (such as drilling depth, rock pressure, rock density, etc.), parameters related to high geological disaster risk can be selected, and data information with no obvious correlation with unfavorable geological bodies can be removed. Third, real-time parameters during drilling (such as drill bit speed, drilling pressure, footage speed, drilling power consumption, etc.) can be monitored to select abnormal fluctuation data, which may indicate that unfavorable geological bodies such as fracture zones and fault mud have been encountered.

[0051] In addition, automated screening methods can be used: two screening ideas are provided here:

[0052] 1. Feature selection based on machine learning: machine learning algorithms such as random forest and Lasso regression can be used to automatically select the most representative and discriminative features. These algorithms can automatically select the optimal feature combination based on the importance of the input data.

[0053] 2. Anomaly detection algorithm: use anomaly detection algorithms (such as Isolation Forest, LOF, etc.) to analyze the input data and select data that may belong to abnormal geological areas, focusing on parts that differ significantly from normal strata.

[0054] In step (3), the data preprocessing is specifically:

[0055] 1. Data cleaning: preliminary cleaning of geological, geophysical and drilling information to remove data points with missing, duplicate or abnormal data to ensure data quality.

[0056] 2. Normalization: For numerical data, normalization methods can be used to convert data of different magnitudes to the same range, eliminating the bias between values, and facilitating subsequent model processing.

[0057] 3. Feature selection: Use algorithms such as correlation analysis, principal component analysis (PCA), etc. to reduce the dimensionality of the data, and select features that have a high contribution to the identification of adverse geological types, reducing the impact of invalid or redundant information on the model.

[0058] It also includes preprocessing for text description data, including but not limited to clarification, word segmentation, etc.

[0059] In step (4), geological data, geophysical inversion imaging information, and drilling information can be considered as three categories of information. Each category of information may include multiple types of information, such as text, image, and numerical information. For different information types, feature extraction of each category of information needs to use multiple deep learning network models for extraction.

[0060] Digital data features can be extracted using DNN models, image data features can be extracted using CNN models, and text description data can be extracted using RNN.

[0061] As a further technical solution:

[0062] (I) Digital data features can be extracted using DNN models, and the specific process is as follows:

[0063] 1. Data preprocessing:

[0064] Normalization: Normalize the input digital data (such as formation depth, rock density, rock thickness, drilling data, etc.) to the value range of [0, 1] or [-1, 1], eliminate the difference between different orders of magnitude, and avoid bias in model training.

[0065] Missing value processing: Use interpolation, mean filling or forward-backward value filling to process missing values in the data, to ensure that the model training does not appear abnormal due to missing data.

[0066] Feature selection: Remove features that have no contribution or low contribution to the identification of adverse geological bodies through correlation analysis or feature importance sorting, to reduce the complexity of the model.

[0067] 2. Input layer:

[0068] The preprocessed digital data is used as the input layer of the DNN (Deep Neural Network), and each feature corresponds to a neuron. The features of digital data include but are not limited to drilling depth, rock pressure, rock density, water content, torque during drilling, etc. which are highly related to the characteristics of adverse geological bodies.

[0069] 3. Hidden layer design:

[0070] Number of layers and number of neurons: The hidden layer usually contains multiple layers, and the number of neurons in each layer is determined according to the complexity of the data and the requirements of the model. Typically, the number of hidden layers can be from 3 to 5 layers, and the number of neurons in each layer decreases (for example, the first layer is 128, the second layer is 64, and the third layer is 32).

[0071] Activation function: The neurons of each hidden layer are activated by the ReLU (Rectified Linear Unit) activation function to introduce non-linear features and speed up model convergence.

[0072] 4. Feature extraction:

[0073] The hidden layer gradually extracts meaningful high-dimensional features by calculating the non-linear relationship between features layer by layer. Finally, high-level abstract features related to adverse geological features are extracted from different depths of the hidden layer, such as the density change trend of a certain geological body and stress distribution anomalies.

[0074] 5. Output layer:

[0075] The final output layer can be a softmax layer for adverse geological body type classification or a linear output layer for predicting the numerical value of the geological parameter. In this way, DNN can effectively process complex numerical features, automatically learn the relationship between numerical features, and extract features useful for identifying adverse geological bodies.

[0076] In the data processing process of the DNN model, the improved way adopted by the embodiment includes:

[0077] 1. Data augmentation: Simulated data based on physical models are used to augment numerical data to improve the generalization ability of the model.

[0078] 2. Regularization techniques: L2 regularization or Dropout techniques are used to prevent DNN model overfitting, especially when the training data is limited. Regularization can significantly improve the robustness of the model.

[0079] (ii) Image data features can be extracted using a CNN model, including but not limited to:

[0080] 1. Data preprocessing:

[0081] Image standardization: Standardize the input geological inversion image, drilling profile image, radar imaging, etc., so that the pixel value is normalized to the range [0, 1], reducing the calculation error.

[0082] Size adjustment: Uniformly adjust different sizes of images to the same size, usually using interpolation or cropping method, so as to batch input into CNN.

[0083] Data Augmentation: By rotating, flipping, scaling, and other operations on images, the diversity of data is increased to avoid model overfitting.

[0084] 2. Input Layer:

[0085] The preprocessed image is input into the input layer of the CNN model, and each pixel of the image serves as an input neuron. Typically, the input data is a three-dimensional tensor (height, width, channel number), such as a two-dimensional profile of geological radar imaging or geophysical imaging results of drilling.

[0086] 3. Convolutional Layer:

[0087] Convolution Operation: In the convolutional layer, multiple convolution kernels (such as 3x3, 5x5 filters) are used to extract local features of the image, such as the boundaries of rock layers, the contours of fractures, and the textures of fissures. The convolution kernel scans the entire image and identifies local features through weighted summation of local information.

[0088] Activation Function: ReLU activation function is usually used after convolution operation to increase nonlinearity and ensure that the network can handle complex features.

[0089] 4. Pooling Layer:

[0090] Max Pooling or Average Pooling: Through pooling operations (such as 2x2 max pooling), the size of the feature map is reduced, important local information is preserved, computational complexity is reduced, and overfitting is avoided. Pooling layers help preserve large-scale information such as rock layer features and fracture trends.

[0091] 5. Fully Connected Layer:

[0092] The output of the convolutional and pooling layers is expanded into a one-dimensional vector and input into the fully connected layer to further extract features related to adverse geological bodies. The fully connected layer learns high-level global features such as overall stratigraphic fault trends and water body distribution through neuron weight learning.

[0093] 6. Output Layer:

[0094] Finally, the image is classified into adverse geological types through softmax or sigmoid functions. For example, the output can be geological body types such as fractures, fracture zones, and caves, or the position and size of adverse geological bodies can be predicted through a regression model.

[0095] In the data processing process of the CNN model, the improved method adopted in this embodiment includes:

[0096] 1. Multi-scale feature extraction: By adding a pyramid pooling layer, geological features at different scales, such as small cracks and large-scale fault structures, can be extracted, improving the recognition ability of complex geological environments.

[0097] 2. Transfer learning: Use pre-trained models (such as ResNet, VGG, etc.) to accelerate the convergence of CNN and improve the feature extraction ability on small data sets.

[0098] (Three) Text description information can be extracted using RNN, including but not limited to RNN, and the specific process is as follows:

[0099] 1. Data preprocessing:

[0100] Text cleaning: Preprocess the text description in the geological exploration report, remove irrelevant characters, punctuation marks, unify terminology, eliminate redundancy and typos, etc.

[0101] Word segmentation: Break down long text data in geological reports into words or phrases, and build a vocabulary. You can use the jieba word segmentation tool for Chinese word segmentation.

[0102] Text encoding: Encode the segmented text into word vectors using Word2Vec, Glove, or pre-trained BERT models, and convert each word into a fixed-dimensional vector representation.

[0103] 2. Input layer:

[0104] Input the encoded text data into the input layer of RNN, and each word or phrase corresponds to a time step of RNN. RNN will process and learn the context relationship between words based on the serialized text information.

[0105] 3. Hidden layer (recurrent layer):

[0106] Recurrent neural network (RNN) or long short-term memory network (LSTM): To better capture long-distance dependencies, it is recommended to use LSTM or GRU instead of traditional RNN. Through the memory unit of LSTM, it can capture the dependence relationship between distant words in geological description, such as the relationship between "fracture zone" and "karst cave" mentioned before and after when describing adverse geological bodies.

[0107] Bidirectional RNN: Through the bidirectional RNN model, information is extracted from the forward and backward directions of the text simultaneously, obtaining a more complete representation of geological features.

[0108] 4. Attention mechanism:

[0109] In the output of the RNN, an attention mechanism is added, allowing the model to focus on important words related to the characteristics of the adverse geological body, such as "fault", "fracture", "water-rich layer", and other key terms, improving the ability to distinguish geological bodies.

[0110] 5. Output layer:

[0111] The final output can be the classification result of the adverse geological body, or a text description of the characteristics of the adverse geological body. By connecting a natural language generation model, an automatic geological report summary can be generated, or a description of the shape, location, etc. of the adverse geological body can be generated.

[0112] In the data processing process of the RNN model, the improved method used in this embodiment includes:

[0113] 1. Pre-training language model: Use BERT, GPT, etc. Pre-training language model as the basis for feature extraction, combined with specially trained geological data set, to enhance the understanding and discrimination ability of the model in the geological field language.

[0114] 2. Multi-task learning: Combine the adverse geological body classification task with the geological description generation task, and improve the understanding ability of the model for text description through multi-task learning.

[0115] This embodiment improves the above-mentioned various model structures, and the specific structure improvements of each model will be described in detail.

[0116] Improvements to DNN structure

[0117] 1.1 Hierarchical design optimization:

[0118] · Adaptive number of layers adjustment: Different geological data complexity may not be consistent. By introducing adaptive network structure, the number of network layers can be dynamically adjusted according to the complexity of the input data, rather than fixed number of layers. Adaptive depth can be achieved by monitoring the change of loss function during training, so that the network has different depths at different stages.

[0119] 1.2 Residual Connections:

[0120] · In deep network, the increase of depth may cause gradient vanishing or gradient explosion problem.

[0121] By adding residual connection (ResNet), it allows some layers to skip and connect directly to deeper layers, ensuring that gradients can be smoothly propagated. This can help DNN avoid performance degradation when processing complex geological data.

[0122] 1.3 Dropout and Batch Normalization:

[0123] • Dropout: Add Dropout operation in each layer to randomly shut down neurons with a certain probability, preventing network overfitting. Especially in geological prediction tasks with less training data, Dropout can effectively improve the generalization ability of the model.

[0124] • Batch Normalization: Perform batch normalization operation on the output of each layer to reduce the internal covariate shift problem, help the model converge faster during training, and avoid local optimal solution.

[0125] 1.4 Mixed activation function:

[0126] • Traditional DNN usually uses a single activation function (such as ReLU). A mixed activation function can be introduced, using different activation functions such as Leaky ReLU, ELU, etc. in different layers to better capture different types of nonlinear relationships. This has good adaptability to complex nonlinear features in geological data.

[0127] In summary, in the DNN model in this embodiment, through adaptive layer adjustment, adding residual connection to ensure that the gradient can propagate smoothly, adding Dropout operation in each layer to effectively improve the generalization ability of the model, performing batch normalization operation on the output of each layer to help the model converge faster during training, and introducing mixed activation function to better adapt to complex nonlinear features in geological data, the traditional DNN model is improved, which is more suitable for processing digital class data in the geological type identification process.

[0128] 2. Improvement of CNN structure

[0129] 2.1 Multi-scale convolution kernel:

[0130] • Multi-scale convolution: Use different size convolution kernels (such as 3x3, 5x5, 7x7) in the convolution layer to capture different scale geological features at the same time. For example, 3x3 convolution kernel can extract detailed information, while 5x5 or 7x7 convolution kernel can help capture large-scale geological structures such as the overall morphology of faults or fracture zones.

[0131] 2.2 Combination of deep convolution and shallow convolution:

[0132] • Feature Pyramid Network (FPN): Fuse high-dimensional features extracted by deep convolution layers and low-dimensional features extracted by shallow convolution layers to extract information from different scales, especially suitable for small cracks and large fault structures existing at the same time in geological images. FPN network can effectively improve the expression ability of features of different scales.

[0133] 2.3 Dilation Convolution:

[0134] • Dilation Convolution can expand the receptive field without increasing the number of parameters, which is particularly suitable for scenes where features are sparse or unevenly distributed in geological images. By increasing the receptive field, Dilation Convolution can better capture long-distance geological features at the same level, improving the model's understanding of complex structures.

[0135] 2.4 Convolution Attention Mechanism:

[0136] • Integrating attention mechanisms into convolutional networks allows the model to focus on key parts of the geological image, such as rock layer boundaries and fault zones, while ignoring irrelevant information. By introducing channel attention mechanisms (such as Squeeze-and-Excitation Networks, SENet) or spatial attention mechanisms, the model's discriminative power can be improved.

[0137] 2.5 Convolution Variants:

[0138] • Depthwise Separable Convolution: This method splits the standard convolution into depthwise convolution and pointwise convolution, significantly reducing the number of parameters and computational complexity, making it suitable for resource-limited application scenarios while maintaining high feature extraction capabilities. This method is suitable for real-time or lightweight geological prediction systems.

[0139] In summary, in the CNN model, through multi-scale convolution, the fusion of high-dimensional features extracted by deep convolution layers and low-dimensional features extracted by shallow convolution layers, the expansion of the receptive field without increasing the number of parameters through Dilation Convolution, the integration of attention mechanisms into convolutional networks, and the use of depthwise separable convolution, the traditional CNN model is improved to better adapt to the processing of image data in the geological type identification process.

[0140] 3. Improvements to RNN structure

[0141] 3.1 Bidirectional LSTM / GRU:

[0142] • Bidirectional RNN: Ordinary RNN or LSTM can only capture one-way sequence information, while bidirectional LSTM / GRU can capture both forward and backward dependencies in geological text. This has obvious advantages for complex language relationships in geological descriptions, such as better understanding of associated words before and after when describing fault structures.

[0143] 3.2 Self-Attention Mechanism:

[0144] • Self-attention mechanisms can help RNNs find important features in long sequence inputs without being limited to relationships between adjacent time steps. With attention mechanisms, RNNs can identify key geological terms and descriptions like "fault zone," "collapse layer," etc. while ignoring secondary information, improving their ability to model long sequence data.

[0145] 3.3 Transformer networks replace RNNs:

[0146] • While RNNs are good at handling sequential data, they have limitations in processing long sequences. Consider replacing traditional RNNs / GRUs / LSTMs with Transformer architectures. Transformers, based on global attention mechanisms, can more efficiently handle long dependencies in geological text and have higher parallelization levels, making them suitable for large-scale geological data processing.

[0147] 3.4 Hierarchical RNNs:

[0148] • Introduce Hierarchical RNN (Hierarchical RNN) architecture for hierarchical processing of geological text. For example, the first layer can handle word relationships within sentences, and the second layer can handle logical relationships between different sentences. This can better capture multi-level information in complex geological reports.

[0149] 3.5 Memory augmentation mechanisms:

[0150] • Improve LSTM or GRU models by introducing memory augmentation mechanisms (Memory Augmented Networks). External memory units record and store important geological information, improving the model's ability to remember long-span information and avoiding information loss. This is very helpful for processing long-term dependency relationships in geological reports, such as the relationship between multiple fault descriptions.

[0151] In summary, in the RNN model, bidirectional RNNs are used to capture both forward and backward dependencies in geological text, self-attention mechanisms are introduced to help RNNs find important features in long sequence inputs, Transformer architectures are used to replace traditional RNNs / GRUs / LSTMs, hierarchical RNN architectures are introduced for hierarchical processing of geological text, and memory augmentation mechanisms are introduced to improve traditional RNN models, making them more suitable for processing image data in the geological type identification process.

[0152] Further, in the step (6), the type identification of the adverse geologic body can be performed using a function including but not limited to a softmax function, and the adverse geologic body morphology description can be performed using a method including but not limited to accessing a natural language generation model for text description, or using a method including but not limited to accessing a generative adversarial network for concept map drawing.

[0153] Further, in the step (7), the loss function used for training the type identification and the morphology description of the adverse geologic body is different.

[0154] The loss function used for training the type identification of the adverse geologic body can use a function including but not limited to a cross-entropy function, wherein a physical constraint can be added:

[0155] L total = L task + L constraint

[0156] wherein L total is the final loss function used; L task is the loss function selected at the beginning of training; and L constraint is a physical constraint term.

[0157] The loss function used for training the morphology description of the adverse geologic body can use a function including but not limited to a mean square error function.

[0158] The two loss functions are weighted to form a joint loss function, so as to realize the overall training of the type identification and the morphology description of the adverse geologic body.

[0159] As shown in FIG. 1, the specific training process of the joint deep learning network of the embodiment includes the following steps: Figure 1

[0160] (1) The geological data, geophysical inversion imaging information and drilling information of the excavated section are screened and summarized, and are arranged into training information. The geophysical inversion imaging information is the geophysical data in the Figure 1 , and the drilling information is the drilling data in the Figure 1 .

[0161] (2) The training information is input into the input layer of the joint deep learning network, and data preprocessing is performed to obtain the geological information, geophysical information and drilling information used in the next step.

[0162] (3) Feature extraction is performed on the three types of information in the extraction layer.

[0163] ​The pre-processed geological information, geophysical information and drilling information are input into the extraction layer, and the characters in the three types of information are extracted by using a deep learning network including but not limited to RNN; the images in the three types of information are extracted by using a deep learning network including but not limited to CNN; and the numbers in the three types of information are extracted by using a deep learning network including but not limited to DNN, so as to obtain geological information features, geophysical information features and drilling information features.

[0164] (4) In the fusion layer, the extracted features in the extraction layer are fused into joint feature representations by using an attention mechanism.

[0165] The geological information features, geophysical information features and drilling information features extracted by the extraction layer are input into the fusion layer, and the features are fused based on the attention mechanism to obtain joint feature representations.

[0166] (5) In the output layer, based on the joint feature representations, different methods are used to complete the type identification and the form description of the adverse geological body.

[0167] Here, the type identification of the adverse geological body includes using a softmax function to identify;

[0168] The form description of the adverse geological body includes using a text description method and a concept map method.

[0169] (6) The above model is trained in a way of minimizing a loss function.

[0170] For type identification, the loss function can be selected to include but not limited to a cross-entropy function; for form description, the loss function can be selected to include but not limited to a mean square error function; the two loss functions are weighted to establish a joint loss function for overall training, and a trained joint deep learning network is obtained.

[0171] The embodiment can realize faster and more accurate type identification and form description of the adverse geological body in front of the tunnel by processing multi-source information through the joint deep learning network.

[0172] Embodiment Two

[0173] The embodiment discloses an adverse geological type identification system based on a joint deep learning network.

[0174] The adverse geological type identification system based on the joint deep learning network comprises:

[0175] A data acquisition module configured to acquire training information including geological data of an excavation section, geophysical inversion imaging information and drilling information;

[0176] The model prediction module is configured to input the training information into a pre-established joint deep learning network model, sequentially perform data preprocessing, feature extraction and feature fusion to obtain a joint feature representation, and based on the joint feature representation, obtain a discriminated type and a form description of the adverse geological body.

[0177] The joint deep learning network model includes a DNN model, a CNN model and an RNN model arranged side by side, and when the features are extracted, the DNN model, the CNN model and the RNN model are used to extract the features of the digital data, the image data and the text description data in the training information, respectively.

[0178] Further, the specific steps performed by the data acquisition module and the model prediction module of the embodiment are as follows:

[0179] (1) Establish an adverse geological body database.

[0180] (2) Screen and aggregate the geological data, geophysical inversion imaging information and drilling information of the excavated section, and organize them into training information.

[0181] (3) Establish a joint deep learning network input layer, and perform data preprocessing on the training information.

[0182] (4) Establish a joint deep learning network extraction layer, and use different deep learning network models to extract features of different types of information in the geological data, geophysical inversion imaging information and drilling information.

[0183] The joint deep learning network model includes a DNN model, a CNN model and an RNN model arranged side by side, and when the features are extracted, the DNN model, the CNN model and the RNN model are used to extract the features of the digital data, the image data and the text description data in the training information, respectively.

[0184] (5) Establish a joint deep learning network fusion layer, and use an attention mechanism to fuse the features of the extracted geological information, geophysical inversion imaging information and drilling information to obtain a joint feature representation.

[0185] (6) Establish a joint deep learning network output layer, and based on the obtained joint feature representation, discriminate the type and describe the form of the adverse geological body.

[0186] (7) Train the above model in a way of minimizing the loss function.

[0187] (8) After the training is completed, input the geological information, geophysical inversion information and drilling information of the forecast section into the joint deep learning network to obtain a prediction result.

[0188] The prediction result output by the joint deep learning network is compared with the result of the adverse geological body database to complete the adverse geological type identification and the shape description of the prediction section.

[0189] The prediction result of the joint deep learning network is compared with the result of the adverse geological body database to prove the reliability of the prediction result of the joint deep learning network.

[0190] Those skilled in the art should understand that the above-mentioned modules or steps of the present application can be realized by a general computer device, and alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device and executed by a computing device, or they can be respectively manufactured into integrated circuit modules, or a plurality of modules or steps thereof can be manufactured into a single integrated circuit module. The present application is not limited to any specific combination of hardware and software.

[0191] Although the specific embodiments of the present application are described above with reference to the accompanying drawings, the present application is not limited to the scope of the drawings. Those skilled in the art should understand that various modifications or variations made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A method for identifying adverse geology types based on a joint deep learning network, characterized in that, The method comprises the following steps: obtaining training information including geological data of a digging section, geophysical inversion imaging information and drilling information; inputting the training information into a pre-established joint deep learning network model, sequentially performing data preprocessing, feature extraction and feature fusion to obtain joint feature representation, and based on the joint feature representation, obtaining the type of the adverse geological body and the morphological description of the adverse geological body; wherein the joint deep learning network model comprises a DNN model, a CNN model and an RNN model arranged side by side, and when performing feature extraction, the DNN model, the CNN model and the RNN model are used to extract features of digital data, image data and text description data in the training information respectively; in the DNN model, the traditional DNN model is improved by self-adaptive layer adjustment, adding residual connection to ensure that the gradient can be smoothly propagated, adding Dropout operation in each layer to effectively improve the generalization ability of the model, performing batch normalization operation on the output of each layer to help the model converge faster during training, and introducing a mixed activation function to better adapt to the complex nonlinear features in the geological data, so as to be more suitable for processing digital data in the geological type identification process; in the CNN model, the traditional CNN model is improved by multi-scale convolution, fusing high-dimensional features extracted by deep convolution layers and low-dimensional features extracted by shallow convolution layers, expanding the receptive field without increasing the number of parameters by using a hollow convolution, integrating the attention mechanism into the convolution network, and using a depth separable convolution, so as to be more suitable for processing image data in the geological type identification process; in the RNN model, the traditional RNN model is improved by using a Transformer architecture to replace the traditional RNN, introducing a hierarchical RNN architecture to process the geological text in layers, and introducing a memory enhancement mechanism, so as to be more suitable for processing image data in the geological type identification process. 2.The method of claim 1, wherein, The joint deep learning network model comprises an input layer, a feature extraction layer, a feature fusion layer and an output layer, wherein: the input layer is used for data preprocessing of the training information; the feature extraction layer is used for feature extraction of digital data, image data and text description data in the preprocessed training information based on the DNN model, the CNN model and the RNN model; the feature fusion layer is used for fusing the features extracted by the feature extraction layer based on the attention mechanism to obtain joint feature representation; the output layer is used for type identification and morphological description of the adverse geological body based on the joint feature representation.

3. The adverse geological type identification method based on the joint deep learning network according to claim 2, wherein: the type of the adverse geological body is identified by using a softmax function, but is not limited to the softmax function; the morphological description of the adverse geological body is described in text by connecting a natural language generation model, and / or is drawn as a concept map by connecting a generative adversarial network. 4.The method of claim 1, wherein, The method also comprises training the joint deep learning network model to obtain the trained joint deep learning network model, wherein: The loss function for training the bad geological body type identification comprises a cross-entropy function with a physical constraint added; The loss function for training the bad geological body shape description comprises a mean square error function; The loss function for bad geological body type identification and the loss function for bad geological body shape description are weighted to form a joint loss function of the joint deep learning network model, and the bad geological body type identification and shape description are trained as a whole. 5.The method of claim 4, wherein, The training of the joint deep learning network model is completed in a manner of minimizing the joint loss function. 6.The method of claim 4, wherein, The method also comprises: establishing a bad geological body database; inputting the geological information, geophysical inversion information and drilling information of the forecast section into the trained joint deep learning network model to obtain a prediction result; comparing the prediction result with the data in the bad geological body database to complete the bad geological type identification and shape description of the forecast section.

7. The method of claim 6, wherein the method further comprises: The bad geological body database is established based on existing bad geological body information, and the data in the bad geological body database covers the type division of bad geology and the feature description of bad geology. 8.The method of claim 1, wherein, The training information comprises at least one of numerical data, image data and text description data. 9.The method of claim 2, wherein, The data preprocessing of the training information comprises normalization processing of the numerical data, and clear processing and word segmentation processing of the text description data.

10. A bad geological type identification system based on a joint deep learning network, characterized in that, The method comprises: a data acquisition module configured to acquire training information, including geological data of a digging section, geophysical inversion imaging information and drilling information; a model prediction module configured to input the training information into a pre-established joint deep learning network model, sequentially perform data preprocessing, feature extraction and feature fusion to obtain joint feature representation, and based on the joint feature representation, obtain the identified type and shape description of the bad geological body; The joint deep learning network model comprises a DNN model, a CNN model and an RNN model arranged side by side, and the DNN model, the CNN model and the RNN model are used to extract the features of the numerical data, the image data and the text description data in the training information during feature extraction; In the DNN model, the traditional DNN model is improved by self-adaptive layer adjustment, addition of residual connection to ensure smooth gradient propagation, addition of a Dropout operation in each layer to effectively improve the generalization ability of the model, batch normalization operation on the output of each layer to help the model converge faster during training, and introduction of a mixed activation function to better adapt to the complex nonlinear features in the geological data, so as to better adapt to the processing of numerical data in the geological type identification process; In the CNN model, the traditional CNN model is improved by multi-scale convolution, fusion of high-dimensional features extracted by deep convolution layers and low-dimensional features extracted by shallow convolution layers, expansion of the receptive field without increasing the parameter amount through atrous convolution, integration of the attention mechanism into the convolution network, and adoption of deep separable convolution, so as to better adapt to the processing of image data in the geological type identification process; In the RNN model, the traditional RNN model is improved by the use of a long short-term memory (LSTM) unit to better adapt to the processing of text description data in the geological type identification process. In the RNN model, the forward and backward dependency information in the geological text is captured simultaneously by the bidirectional RNN, the self-attention mechanism is introduced to help the RNN find important features in long sequence input, the traditional RNN is replaced by the Transformer architecture, the hierarchical RNN architecture is introduced for hierarchical processing of the geological text, and the memory enhancement mechanism is introduced to improve the traditional RNN model, so as to be more suitable for processing of image data in the geological type identification process.

Images