A lithology identification method and system based on the unified characteristics of measured and simulated drilling.
By constructing field and simulated datasets, performing physical feature derivation and feature unification, and utilizing adversarial training and gradient inversion mechanisms, the difficulties in sample acquisition and data mixing in lithology identification in tunnel engineering were solved, achieving high accuracy and robustness in lithology identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2026-05-12
- Publication Date
- 2026-06-09
AI Technical Summary
Existing lithology identification methods in tunnel engineering suffer from problems such as time-consuming and labor-intensive sample acquisition, insufficient training of machine learning models, and difficulty in directly mixing simulated and measured data, resulting in low accuracy and poor adaptability in lithology identification.
By constructing field datasets and simulated datasets, physical features are derived and preprocessed. Combining one-dimensional convolutional neural networks and bidirectional long short-term memory networks, and utilizing adversarial training and gradient inversion mechanisms, feature unification between simulated drilling data and measured drilling data is achieved, eliminating distribution differences and improving the accuracy and robustness of the model.
Accuracy and robustness of lithology identification were achieved with limited measured data, overcoming the bottleneck of traditional methods that rely on a large amount of field-labeled data, and improving the model's adaptability and generalization ability.
Smart Images
Figure CN122174028A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent lithology identification technology, and in particular relates to a lithology identification method and system based on the unified characteristics of measured simulated drilling. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Lithology identification is a fundamental task in geotechnical engineering fields such as tunnels, water conservancy, and slope protection. Accurate lithology identification provides a scientific basis for construction scheme design, geological hazard risk assessment, and project feasibility evaluation. Currently, the commonly used lithology identification method is core drilling. Experts can identify lithology by visually inspecting the drilled core or using laboratory methods such as electron microscopy. This method has the advantages of being intuitive and reliable, but identifying lithology through core drilling is time-consuming, labor-intensive, and costly.
[0004] Measurement while drilling (MSWL) technology can store drilling parameters such as drilling speed, drilling pressure, and drilling torque in real time. These parameters contain rich geological information and can provide valuable data support for lithology identification. Currently, numerous studies have used drilling parameters as data-driven methods for lithology identification; however, the following technical challenges still exist: (1) Using machine learning models for lithology identification has a high accuracy rate, but the accuracy of such machine learning methods depends on the size and quality of the sample. However, obtaining lithology labels at the tunnel site is time-consuming and laborious. Models built with only a small amount of field measurement data will lead to problems such as insufficient training and overfitting. (2) A large amount of simulation data can be generated for training by drilling numerical simulation. However, there are many differences between simulation data and field measured data in terms of geological complexity and noise level, making it difficult to directly mix simulation data with measured data to improve model performance. Summary of the Invention
[0005] To overcome the shortcomings of the prior art, the present invention provides a lithology identification method and system based on the unified characteristics of measured and simulated drilling, which eliminates the distribution difference between simulated drilling data and measured drilling data, and can achieve accurate identification of lithology when the sample size in the field is small.
[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a lithology identification method based on the unified characteristics of measured simulated drilling.
[0007] A lithology identification method based on unified measured simulated drilling characteristics includes: Construct field datasets and simulation datasets; Preprocessing is performed on the established field dataset and simulation dataset; physical features are derived from the drilling parameters in the two types of datasets after preprocessing, and the derived physical features are combined with the basic drilling parameters to form new field dataset and simulation dataset; The newly spliced field dataset and simulation dataset are subjected to two-stage feature extraction, and the hidden features are obtained by unfolding using a fully connected layer. The obtained hidden features are input into the label classifier and the domain discriminator to calculate the classification loss and adversarial loss respectively; based on the obtained classification loss and adversarial loss, the model is trained using backpropagation and optimizer iteration; the performance of the trained model is verified. Based on the performance-verified model, the drilling characteristics were unified through actual measurement simulation to obtain common features; the lithology of the unexcavated section was identified using drilling parameters.
[0008] Furthermore, physical features are derived from the drilling parameters in the two preprocessed datasets, including: based on the four fundamental parameters of drill pressure, torque, rotational speed, and drilling speed, derivative physical quantities including mechanical specific energy, drill pressure-to-drill speed ratio, and torque-to-drill pressure ratio are calculated; wherein, the mechanical specific energy is expressed as... ; in, Indicates mechanical specific energy; Indicates drilling pressure. Indicates rotational speed. Indicates torque, Indicates drilling speed; This represents the cross-sectional area of the drill bit, used to characterize the energy required to break a unit volume of rock.
[0009] Furthermore, the domain discriminator incorporates a domain classification uncertainty penalty, which adjusts the degree of influence of the domain classification uncertainty penalty by the entropy of the output probability of the domain discriminator; and the domain classification uncertainty penalty takes effect only during backpropagation.
[0010] Furthermore, the field dataset is obtained by collecting drilling parameters through advanced drilling experiments in tunnel engineering sites and combining them with lithology identification in the core drilling laboratory; the simulation dataset is obtained by generating corresponding drilling parameters through numerical simulation of drilling experiments under different lithological conditions.
[0011] Furthermore, the two-stage feature extraction includes: first, extracting local features of drilling parameters; and then, extracting bidirectional contextual information and long-term dependencies of drilling data.
[0012] Furthermore, the obtained hidden features are input into the label classifier to obtain the classification probability, and the classification loss is obtained by calculating the cross-entropy loss function between the predicted and actual labels; the obtained hidden features are input into the neighborhood discriminator to obtain the neighborhood discrimination probability, and the adversarial loss is obtained by calculating the cross-entropy loss function between the predicted and actual labels.
[0013] Furthermore, the backpropagation includes: first, fixing the neighborhood discriminator, updating the feature extractor and label classifier to minimize the classification loss; then, fixing the feature extractor, updating the neighborhood discriminator to minimize the adversarial loss, and updating all network parameters simultaneously through one forward-backward propagation.
[0014] A second aspect of the present invention provides a lithology identification system based on the unified characteristics of measured simulated drilling.
[0015] A lithology identification system based on unified measured simulated drilling characteristics includes: The dataset building module is configured to build both field datasets and simulation datasets. The data preprocessing and physical feature derivation module is configured to: perform preprocessing operations on the established field dataset and simulation dataset; perform physical feature derivation on the drilling parameters in the two types of datasets after preprocessing; and concatenate the derived physical parameters with the basic drilling parameters to form new field dataset and simulation dataset. The feature extraction module is configured to perform two-stage feature extraction on the newly spliced field dataset and simulation dataset, and use fully connected layers to expand and obtain hidden features. The model training module is configured to: input the obtained hidden features into the label classifier and the domain discriminator to calculate the classification loss and adversarial loss respectively; and train the model using backpropagation and optimizer iteration based on the obtained classification loss and adversarial loss. The model validation module is configured to perform performance validation on the trained model. The lithology identification module is configured to: unify the measured simulated drilling characteristics based on the performance-verified model to obtain common features; and identify the lithology of the unexcavated section using drilling parameters.
[0016] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a lithology identification method based on the uniformity of measured simulated drilling characteristics as described in the first aspect of the present invention.
[0017] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the lithology identification method based on the unified measured simulated drilling characteristics as described in the first aspect of the present invention.
[0018] The above one or more technical solutions have the following beneficial effects: (1) This invention derives physical features from drilling parameters in the preprocessed field dataset and simulation dataset, and concatenates the derived physical features with the basic drilling parameters to form new field datasets and simulation datasets, which are then used as the measured drilling data and simulated drilling data to be analyzed subsequently. Local features and bidirectional contextual information of the drilling data are extracted by fusing a one-dimensional convolutional neural network and a bidirectional long short-term memory network. Combined with an adversarial training mechanism and a gradient inversion layer, the data features of the simulated drilling data and the measured drilling data are unified. Through the adversarial training strategy and gradient inversion mechanism, the massive amount of simulated data generated by numerical simulation and the small amount of measured data are unified in terms of features, effectively expanding the size of the field dataset and reducing the cost of collecting a large amount of field dataset. By fusing the low-cost generation capability of simulated data with the real geological features of measured data, the bottleneck problem of traditional methods relying on a large amount of field-labeled data is solved, effectively alleviating the problems of insufficient model training, poor generalization ability, and overfitting in small sample scenarios, enabling the lithology identification model to maintain high accuracy and robustness even with a small amount of measured data.
[0019] (2) This invention uses a joint feature extraction module of a one-dimensional convolutional neural network and a bidirectional long short-term memory network, combined with a domain discriminator in adversarial training, to force the feature extractor to generate domain-invariant features, eliminating the distribution differences between simulated and measured data in terms of noise level, geological complexity, and dynamic response. By introducing a domain classification uncertainty penalty into the domain discriminator, it encourages the domain discriminator to "hesitate" on intermediate features, making it more adversarial than existing DANN models and achieving more thorough feature confusion. By using a gradient reversal mechanism to optimize the feature extraction process in reverse, the two types of data are aligned in the feature space, thereby breaking through the direct mixing limitation of simulated and measured data and improving the model's adaptability to actual engineering scenarios.
[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0022] Figure 1 This is a flowchart of a lithology identification method based on the unified characteristics of measured simulated drilling, as described in Embodiment 1 of the present invention.
[0023] Figure 2 This is a schematic diagram of the forward propagation and loss function calculation process in Embodiment 1 of the present invention.
[0024] Figure 3 This is a schematic diagram of the model dimensions in Embodiment 1 of the present invention; wherein, Figure 3 In the diagram, (a) and (b) represent the drilling model from the frontal and lateral viewpoints, respectively. Figure 3 In the diagram, (c) and (d) represent the drill bit model from the frontal and lateral viewpoints, respectively.
[0025] Figure 4 This is a schematic diagram of the 1DCNN structure in Embodiment 1 of the present invention.
[0026] Figure 5 This is a schematic diagram of the BiLSTM structure in Embodiment 1 of the present invention.
[0027] Figure 6 This is a schematic diagram of the experimental results of the lithology identification model based on the unified characteristics of measured-simulated drilling data in Embodiment 1 of the present invention; wherein, Figure 6 (a) in the figure is a schematic diagram of the experimental results of the training set. Figure 6 (b) in the diagram is a schematic diagram of the test results for the test set.
[0028] Figure 7 This is a schematic diagram of the actual lithology identification results in Embodiment 1 of the present invention. Detailed Implementation
[0029] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0030] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.
[0031] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0032] The overall approach proposed in this invention is as follows: This invention provides a lithology identification method based on the unification of measured and simulated drilling features. Physical features are derived from drilling parameters in two preprocessed datasets, and the derived physical features are concatenated with the basic drilling parameters to form new field and simulated datasets. Through adversarial training strategies and gradient inversion mechanisms, the feature extractor is forced to generate domain-invariant features, eliminating the distribution differences between simulated and measured drilling data, thus unifying the features of measured and simulated drilling data and achieving accurate lithology identification with small sample sizes.
[0033] Example 1 This embodiment discloses a lithology identification method based on the unified characteristics of measured simulated drilling.
[0034] like Figure 1 As shown, a lithology identification method based on unified measured simulated drilling characteristics includes: Step S1: Construct the field dataset and the simulation dataset; Step S2: Perform preprocessing operations on the established field dataset and simulation dataset; derive physical features from the drilling parameters in the two types of datasets after preprocessing, and combine the derived physical parameters with the basic drilling parameters to form a new field dataset and simulation dataset; Step S3: Perform two-stage feature extraction on the newly spliced field dataset and simulation dataset, and use a fully connected layer to expand and obtain hidden features; Step S4: Input the obtained hidden features into the label classifier and the domain discriminator to calculate the classification loss and adversarial loss respectively; based on the obtained classification loss and adversarial loss, use backpropagation and optimizer iteration to train the model; Step S5: Validate the performance of the trained model; Step S6: Based on the performance-verified model, perform real-world simulation drilling feature unification to obtain common features; use drilling parameters to identify the lithology of the unexcavated section.
[0035] Based on the above process, this invention eliminates the distribution discrepancy between simulated drilling data and measured drilling data, enabling accurate identification of lithology even with a small sample size in the field. To facilitate understanding of the technical solution of this invention, the specific implementation steps are further explained and described below.
[0036] Step S1: Construct the field dataset and the simulation dataset.
[0037] Advanced drilling experiments were conducted at the tunnel engineering site to collect drilling parameters, and core samples were collected for laboratory lithology identification to construct a field dataset. Numerical simulations were carried out, setting up drilling experiments with different lithologies and recording the corresponding drilling parameters to construct a simulation dataset.
[0038] As an optional embodiment, particle flow simulation software can be used to simulate drilling experiments with different lithologies. The rock mass drilling model consists of a rock mass model and a drill bit model. In the actual implementation of this embodiment: First, a rock mass model is established, using a layered modeling approach, with particle size gradually increasing from the inside out; the innermost layer is the part where the drill bit directly contacts the rock particles.
[0039] Subsequently, a drill bit model was created. The PDC drill bit was treated as a rigid body that does not wear. The `wall` command was used to model the drill bit, generating eight cutting teeth evenly distributed around it. The drilling model and drill bit model are shown below. Figure 3 As shown; Figure 3 In the diagram, (a) and (b) represent the drilling model from the frontal and lateral viewpoints, respectively. The blue area represents the outer layer of the model, the red area represents the middle layer of the model, and the green area represents the inner layer of the model. Figure 3 In the diagram, (c) and (d) represent the drill bit model from the frontal and lateral perspectives, respectively, and the drilling direction is marked in the lateral perspective.
[0040] In addition, before conducting numerical simulations of rock drilling, the basic mechanical parameters of different lithologies (mainly those included in the field dataset) were first determined in the laboratory using basic mechanical instruments, such as uniaxial compressive strength, tensile strength, and shear strength. Then, the model's micro-parameters were adjusted for micro-calibration to ensure that the numerical simulation results were consistent with the actual mechanical test results. After that, simulated drilling experiments were conducted on different lithologies, and the drilling pressure, drilling speed, torque, rotational speed, and corresponding lithology labels during the simulated drilling process were collected as the simulation dataset.
[0041] Step S2: Preprocess the established field dataset and simulation dataset; derive physical features from the drilling parameters in the two datasets after preprocessing, and concatenate the derived physical features with the basic drilling parameters to form new field datasets and simulation datasets. The preprocessing operations include data denoising, equidistant resampling alignment, standardization, and dataset partitioning.
[0042] Data noise reduction: Due to the stability of drilling rig equipment and the sampling accuracy of drilling parameter acquisition equipment, the acquired drilling data will exhibit irregularities and roughness due to some instantaneous changes. Therefore, wavelet transform or adaptive filtering is used to smooth the data, reduce noise, minimize curve fluctuations, and preserve the trend and characteristics as much as possible for both measured and simulated drilling data.
[0043] Equidistant resampling: Simulated drilling data has an extremely high sampling frequency due to stable drilling rate control; however, measured drilling data has a larger sampling interval due to limitations in sensor accuracy, drilling efficiency, or storage. This difference in sampling density makes it impossible to extract features from both measured and simulated drilling data simultaneously. Therefore, equidistant resampling alignment is required before feature extraction. This involves downsampling the simulated drilling data and linearly interpolating the measured drilling data to ensure that the number of sampling points in the simulated and measured drilling data is consistent across the same drilling distance.
[0044] Standardization: To reduce the numerical differences between measured and simulated drilling data and improve the practicality and accuracy of the model, the data is standardized. Specifically, max-min standardization is used to scale the two types of data (simulated and measured drilling data) to the range of [-1, 1], preserving the relative size of the original data without changing the distribution characteristics of the data.
[0045] Dataset partitioning: To ensure the independence between the training and test sets, the actual drilling data were set as the training and test sets in an 8:2 ratio. The purpose of this invention is to improve the accuracy and generalization of a lithology identification model built from a small amount of actual drilling data using a large amount of simulated drilling data; therefore, all simulated drilling data were set as the training set.
[0046] Physical features are derived from the drilling parameters in the preprocessed field dataset and simulation dataset. Specifically, based on the four fundamental parameters of drill pressure, torque, rotational speed, and drilling speed, derivative physical quantities including mechanical specific energy, drill pressure-to-drill speed ratio, and torque-to-drill pressure ratio are calculated. The mechanical specific energy is expressed as: ; in, Indicates mechanical specific energy; Indicates drilling pressure. Indicates rotational speed. Indicates torque, Indicates drilling speed; This represents the cross-sectional area of the drill bit, used to characterize the energy required to break a unit volume of rock.
[0047] The three derivative parameters mentioned above are concatenated with the original four basic drilling parameters to form a seven-dimensional input feature vector, which is then input into the two-stage feature extraction module.
[0048] Step S3: Perform two-stage feature extraction on the newly spliced field dataset and simulation dataset, and use a fully connected layer to expand and obtain hidden features.
[0049] Using the stitched field dataset and simulated dataset as the measured and simulated drilling data to be analyzed, respectively, these two types of data transform the original four drilling parameters into a seven-dimensional feature vector: the original drill pressure, torque, rotational speed, and drilling speed, as well as the derived mechanical energy, drill pressure-to-speed ratio, and torque-to-drill pressure ratio. Based on this, local features of the drilling parameters are first extracted, followed by the extraction of bidirectional contextual information and long-term dependencies from the drilling data. Finally, hidden features are obtained through fully connected layer expansion. , This serves as input features for subsequent domain discrimination and lithology identification. It's important to note that during model training, the fully connected layers, when expanded, yield hidden features; after model training is complete, the expanded layers yield common features. Feature extraction can utilize a combination of one-dimensional convolutional neural networks and bidirectional long short-term memory networks.
[0050] Step S3-1: Extract local features.
[0051] Drilling data has typical time series characteristics. 1DCNN is good at processing time series data and has a strong ability to extract local features. Therefore, 1DCNN can be used to extract features from drilling data.
[0052] like Figure 4 As shown, 1DCNN mainly consists of an input layer, convolutional layers, pooling layers, activation layers, and fully connected layers. Among them: 1) Input layer: Receives time series data as input to the model. This data is represented as a two-dimensional matrix, where a row represents a sequence and a column represents an element in the sequence.
[0053] 2) Convolutional Layers: This is the core of 1D CNNs. It uses one-dimensional convolutional kernels to perform convolution operations on the input sequence to extract local features. The convolutional kernel slides across the one-dimensional input data, capturing local patterns and features by performing weighted summations and other operations with data at corresponding positions. The size (width) and number of convolutional kernels are important parameters, determining the granularity of feature extraction and the complexity of the model. Smaller convolutional kernels capture more detailed local features, while increasing the number of kernels can extract more different types of features.
[0054] 3) Activation Layer: The activation layer, or activation function, is typically used after the convolutional layer to introduce nonlinearity, enabling the model to learn the complex nonlinear mappings inherent in the data. Common activation functions include Sigmoid, Tanh, ReLU, and LeakyReLU, each suitable for different types of data features and model structures. Among them, the ReLU activation function is widely used in convolutional neural networks because it preserves the gradient along the positive half-axis.
[0055] 4) Pooling Layer: Its function is to downsample the data, reduce data dimensionality, reduce computational cost, and prevent overfitting to some extent while preserving the main features of the data. Common pooling operations include max pooling and average pooling. In the one-dimensional case, the pooling window slides along the sequence dimension.
[0056] 5) Fully connected layer: This layer aggregates the features extracted and processed by the previous layers, mapping the high-level features to a global category or label, thereby achieving tasks such as classification or regression. Each neuron in a fully connected layer is connected to all neurons in the previous layer.
[0057] Step S3-2: Extract bidirectional contextual information and long-term dependencies from the drilling data.
[0058] like Figure 5 As shown, BiLSTM consists of two independent LSTMs. The forward LSTM processes the input data in the original sequence order, learning the parameter changes before each time step; the backward LSTM processes the data in reverse order, capturing the parameter influence after each time step. Both have the same hyperparameters but independent weights to ensure the learning of features in different directions. After all input sequence data has been computed, the outputs of the forward and backward LSTMs are fused, concatenating the outputs along the feature dimension. This results in the final features containing both past and future information, i.e., bidirectional contextual information and long-term dependencies.
[0059] For the drilling sequence data processed in this paper, 1DCNN is first used to extract local features, and then BiLSTM is introduced to analyze global temporal features and long-term dependencies. This can fuse multi-scale features and accurately characterize the drilling state features.
[0060] Step S3-3: Obtain the hidden features by unfolding the fully connected layer.
[0061] After the feature extractor, multi-scale drilling features are generated. For the same lithology, the measured drilling data and simulated drilling data have different drilling features due to distribution differences. After adversarial training and gradient inversion, the hidden features are generated. These hidden features refer to the common characteristics of the measured and simulated drilling data. To better understand how the fully connected layers work, the network structure is shown below: This paper presents a partial structure and parameters of the feature extractor for a lithology identification model based on 1DCNN-BiLSTM-DANN. B The batch size is shown in Table 1. Table 1 Feature Extractor Network Structure Parameters
[0062] The structure and parameters of the lithology identification model classifier based on 1DCNN-BiLSTM-DANN are presented in Table 2. Table 2. Classifier Network Structure Parameters
[0063] When the category classifier is used as a domain discriminator, the output dimension parameter of the last fully connected layer of the category classifier is 1; when the category classifier is used as a label classifier, the output dimension parameter is the category to be classified for lithology identification. Here, the parameters of the domain discriminator are used for illustration.
[0064] Step S4: Input the obtained hidden features into the label classifier and the domain discriminator to calculate the classification loss and adversarial loss respectively; based on the obtained classification loss and adversarial loss, use backpropagation and optimizer iteration to train the model.
[0065] Step S4-1: Forward propagation and loss function calculation.
[0066] The obtained hidden features The predicted classification probabilities are obtained by inputting the data into a label classifier. The classification loss is then calculated by performing a cross-entropy loss function between the predicted and actual labels. The resulting hidden features are then processed... The neighborhood discrimination probability is obtained by inputting the data into the neighborhood discriminator. The adversarial loss is obtained by calculating the cross-entropy loss function between the predicted neighborhood label and the actual label. The sum of the obtained classification loss and adversarial loss is used as the total loss.
[0067] The cross-entropy loss function of the label classifier can be expressed as: ; in, y For lithological category labels, For label classifiers.
[0068] The cross-entropy loss function of the neighborhood discriminator can be expressed as: ; in, d For domain category labels, This is a domain discriminator.
[0069] The formula for calculating the total loss can be expressed as: ; in, It is a hyperparameter used to control the weights of the neighborhood discriminator when performing a reverse update of the cross-entropy loss function.
[0070] Step S4-2: Backpropagation and parameter update.
[0071] like Figure 2 As shown, backpropagation and optimizer iterations are used to continuously improve model performance and update network parameters. Specifically, backpropagation includes: first, fixing the neighborhood discriminator, updating the feature extractor and label classifier to minimize classification loss; then, fixing the feature extractor, updating the neighborhood discriminator to minimize adversarial loss, and updating all network parameters simultaneously through one forward-backward propagation. The feature extractor, label classifier, and neighborhood discriminator are specifically: 1) Feature Extractor Used to extract common features from measured drilling data and simulated drilling data, mapping them to the same feature space for subsequent training.
[0072] 2) Tag Classifier The main purpose is to achieve the classification task of measured drilling data and simulated drilling data, and through continuous training, improve the feature extractor to extract deeper category features.
[0073] 3) Domain Discriminator The main purpose is to distinguish whether the input samples come from actual drilling data and simulated drilling data. By using the gradient inversion layer (GRL) to form an adversarial training with the feature extractor, the feature extractor is forced to learn features that are useful for classification but irrelevant to domain discrimination, thereby achieving feature alignment between actual drilling data and simulated drilling data.
[0074] Based on this, the training of DANN involves: first, performing forward propagation, where measured drilling data and simulated drilling data are processed by a feature extractor. Obtain features Next, the obtained features It will be simultaneously passed to the label classifier and Domain Discriminator Used to predict category labels and domain tags Among them, the category label is the actual label of the data, and the domain label includes the source domain and the target domain, which are usually represented by 0 or 1.
[0075] During backpropagation, the label classifier and the neighborhood discriminator use the same cross-entropy loss function as in the forward propagation process for loss calculation. The core idea of DANN is to use a gradient inversion layer (GRL) during backpropagation, which automatically inverts the gradient when the neighborhood loss occurs, forcing the feature extractor to learn neighborhood-independent features. The definition of gradient inversion is as follows: ; Furthermore, this invention introduces a domain classification uncertainty penalty into the domain discriminator. This penalty adjusts its influence by using the entropy of the discriminator's output probability; and it only applies during backpropagation. Therefore, after backpropagation, the total loss of the DANN is expressed as: ; in, and These represent hyperparameters, used to control the weights of the neighborhood discriminator when it performs reverse updates of the cross-entropy loss function and the entropy of the output probability, respectively. Entropy, representing the output probability of the domain discriminator, encourages the domain discriminator to "hesitate" about intermediate features (i.e., the output domain probability is close to 0.5, thus eliminating the discriminability between simulated and measured data in the feature space to a greater extent). This is more adversarial than the standard DANN and can achieve more thorough feature confusion.
[0076] The goal of the feature extractor and label classifier is to minimize the class classification loss, while the goal of the domain discriminator is to maximize the domain discrimination loss through the gradient inversion (GRL) mechanism. The overall goal of the DANN is to learn unified features that can perform effective classification and have strong generalization ability across different domains. Ideally, through adversarial training, the feature extractor can not only unify the features of real-world and simulated drilling data, but also perform effective classification learning on the label classifier. In this case, the domain discriminator outputs a discrimination probability of 1 / 2, meaning that the domain discriminator cannot determine whether the input features come from the source domain or the target domain.
[0077] The model's overall input consists of a simulated drilling dataset and a measured drilling dataset, including simulated drilling parameters and their corresponding lithology labels, and measured drilling parameters and their corresponding lithology labels. The output consists of two parts: the first part is a label classifier, which outputs the lithology labels, and the second part is a neighborhood discriminator, which outputs either the simulated drilling data or the measured drilling data.
[0078] Step S5: Validate the performance of the trained model.
[0079] This model refers to a lithology identification model based on unified features of simulated drilling data from 1DCNN-BiLSTM-DANN. 1DCNN-BiLSTM is a feature extractor used to extract multi-scale drilling features from simulated and measured drilling data. The role of DANN is adversarial training, which ensures the lithology identification effect while reducing the distribution difference between simulated and measured drilling data, so that the features extracted by the feature extractor are common features of simulated and measured drilling data.
[0080] After reaching the specified number of training iterations, the model training is complete. Subsequently, the trained model's performance is validated by inputting the test set into the model and calculating the lithology identification effect. As an optional implementation, the confusion matrix, accuracy (A), precision (P), recall (R), and F1 score can be used to validate the model's performance.
[0081] like Figure 6 The figure shows the experimental results of a lithology identification model based on the unified characteristics of measured and simulated drilling data; among which, Figure 6 (a) in the figure is a schematic diagram of the experimental results of the training set. Figure 6 (b) in the diagram shows the experimental results for the test set. Figure 6 As shown in (a), the diagonal elements of the confusion matrix in the training set are significantly higher than the off-diagonal elements. The accuracy, precision, recall, and F1 score of the training set are 94.9%, 95.4%, 94.8%, and 95.1%, respectively, indicating that the model has a good learning ability for the complex nonlinear relationship between drilling parameters and lithology in the known data. Figure 6 As can be seen in (b), the main diagonal elements of the confusion matrix in the test set still dominate. For example, 11 items in category "B" were correctly predicted, and only 1 item was misclassified as "C". Although the performance is slightly inferior to that of the training set, the evaluation indicators of the test set are still at a high level, indicating that the model has good generalization ability and can effectively identify various types of rock masses.
[0082] Step S6: Based on the performance-validated model, use drilling parameters to identify the lithology of the unexcavated section.
[0083] By inputting the drilling parameters of the rock to be identified into the performance-verified model, the lithology of the rock can be identified. In this embodiment, drilling parameter data for a tunnel X5DK0+753~+759.4, totaling 6.4m, was tracked and acquired, including drilling speed, torque, rotational speed, and drilling pressure. The tunnel was divided into sections of 0.2m each, and rock samples were taken from the vicinity of the borehole during the later stages of drilling. Laboratory tests were conducted to obtain the true lithology labels. The tests showed that this section of lithology contained only three types: A, B, and C.
[0084] The drilling parameters obtained during the drilling process are segmented and input into the lithology identification model. The lithology results predicted by the model are compared with the actual lithology labels. Figure 6 As shown. From Figure 7 It is easy to see that the lithology identification results of the model of the present invention are basically consistent with the actual lithology labels. Only in groups 18 and 30 were there misjudgments in the prediction results. The prediction accuracy rate is over 93%, which shows that the model of the present invention has a high lithology identification accuracy and good engineering adaptability.
[0085] Example 2 This embodiment discloses a lithology identification system based on the unified characteristics of measured simulated drilling.
[0086] A lithology identification system based on unified measured simulated drilling characteristics includes: The dataset building module is configured to build both field datasets and simulation datasets. The data preprocessing and physical feature derivation module is configured to: perform preprocessing operations on the established field dataset and simulation dataset; perform physical feature derivation on the drilling parameters in the two types of datasets after preprocessing; and concatenate the derived physical parameters with the basic drilling parameters to form new field dataset and simulation dataset. The feature extraction module is configured to perform two-stage feature extraction on the newly spliced field dataset and simulation dataset, and use fully connected layers to expand and obtain hidden features. The model training module is configured to: input the obtained hidden features into the label classifier and the domain discriminator to calculate the classification loss and adversarial loss respectively; and train the model using backpropagation and optimizer iteration based on the obtained classification loss and adversarial loss. The model validation module is configured to perform performance validation on the trained model. The lithology identification module is configured to: unify the measured simulated drilling characteristics based on the performance-verified model to obtain common features; and identify the lithology of the unexcavated section using drilling parameters.
[0087] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.
[0088] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a lithology identification method based on the uniformity of measured simulated drilling characteristics as described in Embodiment 1 of this disclosure.
[0089] Example 4 The purpose of this embodiment is to provide an electronic device.
[0090] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in a lithology identification method based on the uniformity of measured simulated drilling characteristics as described in Embodiment 1 of this disclosure.
[0091] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.
[0092] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.
[0093] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A lithology identification method based on measured analog drilling feature uniformity, characterized in that, include: Construct field datasets and simulation datasets; Preprocess the established field dataset and simulation dataset; Physical features are derived from the drilling parameters in the two types of preprocessed datasets, and the derived physical features are combined with the basic drilling parameters to form new field datasets and simulation datasets. The newly spliced field dataset and simulation dataset are subjected to two-stage feature extraction, and the hidden features are obtained by unfolding using a fully connected layer. The obtained hidden features are input into the label classifier and the domain discriminator to calculate the classification loss and adversarial loss respectively; based on the obtained classification loss and adversarial loss, the model is trained using backpropagation and optimizer iteration; the performance of the trained model is verified. Based on the performance-verified model, we conducted experimental simulations to unify drilling characteristics in order to obtain common features. Use drilling parameters to identify the lithology of unexcavated sections.
2. The lithology identification method based on the unified characteristics of measured simulated drilling as described in claim 1, characterized in that, The physical characteristics are derived from the drilling parameters in the two types of data sets after preprocessing, including: based on the four basic parameters of drilling pressure, torque, rotating speed and drilling speed, the derivative physical quantities including mechanical specific energy, drilling pressure-drilling speed ratio, and torque-drilling pressure ratio are calculated; wherein the mechanical specific energy is expressed as ; in, Indicates mechanical specific energy; Indicates drilling pressure. Indicates rotational speed. Indicates torque, Indicates drilling speed; This represents the cross-sectional area of the drill bit, used to characterize the energy required to break a unit volume of rock.
3. The lithology identification method based on the unified characteristics of measured simulated drilling as described in claim 1, characterized in that, The domain discriminator incorporates a domain classification uncertainty penalty, the influence of which is adjusted by the entropy of the output probability of the domain discriminator; and the domain classification uncertainty penalty takes effect only during backpropagation.
4. The lithology identification method based on the unified characteristics of measured simulated drilling as described in claim 1, characterized in that, The field dataset was obtained by collecting drilling parameters through advanced drilling experiments in tunnel engineering sites and combining them with lithology identification in the core sampling laboratory; the simulation dataset was obtained by generating corresponding drilling parameters through numerical simulation of drilling experiments under different lithological conditions.
5. The lithology identification method based on the unified characteristics of measured simulated drilling as described in claim 1, characterized in that, The two-stage feature extraction includes: first, extracting local features of drilling parameters; and then, extracting bidirectional contextual information and long-term dependencies of drilling data.
6. The lithology identification method based on the unified characteristics of measured simulated drilling as described in claim 1, characterized in that, The obtained hidden features are input into the label classifier to obtain the classification probability, and the classification loss is obtained by calculating the cross-entropy loss function between the predicted and actual labels. The obtained hidden features are input into the neighborhood discriminator to obtain the neighborhood discrimination probability, and the adversarial loss is obtained by calculating the cross-entropy loss function between the predicted and actual labels.
7. The lithology identification method based on the unified characteristics of measured simulated drilling as described in claim 1, characterized in that, The backpropagation includes: first, fixing the neighborhood discriminator, updating the feature extractor and label classifier to minimize the classification loss; then, fixing the feature extractor, updating the neighborhood discriminator to minimize the adversarial loss, and updating all network parameters simultaneously through one forward-backward propagation.
8. A lithology identification system based on unified measured simulated drilling characteristics, characterized in that, include: The dataset building module is configured to build both field datasets and simulation datasets. The data preprocessing and physical feature derivation module is configured to perform preprocessing operations on the established field dataset and simulation dataset. Physical features are derived from the drilling parameters in the two types of preprocessed datasets, and the derived physical features are combined with the basic drilling parameters to form new field datasets and simulation datasets. The feature extraction module is configured to perform two-stage feature extraction on the newly spliced field dataset and simulation dataset, and use fully connected layers to expand and obtain hidden features. The model training module is configured to: input the obtained hidden features into the label classifier and the domain discriminator to calculate the classification loss and adversarial loss respectively; and train the model using backpropagation and optimizer iteration based on the obtained classification loss and adversarial loss. The model validation module is configured to perform performance validation on the trained model. The lithology identification module is configured to: unify the measured simulated drilling characteristics based on the performance-verified model in order to obtain common features; Use drilling parameters to identify the lithology of unexcavated sections.
9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps in the lithology identification method based on the unified characteristics of measured simulated drilling as described in any one of claims 1-7.
10. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the lithology identification method based on the unified characteristics of measured simulated drilling as described in any one of claims 1-7.