A method for intelligent identification of poor geological bodies while drilling based on multi-modal fusion

Abstract

The application provides a method for intelligent identification of bad geological bodies while drilling based on multi-modal fusion, which comprises data collection, data preprocessing, feature extraction, modal common and private feature extraction, modal fusion and bad geological body identification. The application collects four modal data of drilling rig operation parameters, drilling vibration parameters, drilling acoustic signals and drilling resistivity imaging to form a multi-modal drilling data set. The multi-modal data can make full use of the complementarity of different data sources, provide more comprehensive analysis for the identification of bad geological bodies, improve the accuracy and reliability of identification, enhance the anti-noise ability through cross-validation of different data modalities, and filter out environmental interference signals. In addition, multi-modal fusion can adapt to more complex geological environments, especially in the case of coexistence of multiple geological anomalies, can better identify potential bad geological bodies, provide real-time and accurate feedback, and provide technical support for complex geological conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bad geological body identification, and particularly relates to a drilling bad geological body intelligent identification method based on multi-modal fusion. BACKGROUND

[0002] With the rapid development of modern infrastructure construction, underground engineering plays an increasingly important role in the fields of transportation, energy development, and urban underground space utilization. However, due to the complexity of underground geological conditions, various bad geological bodies such as faults, karst caves, soft interlayers, karst, expansive soils, and areas with abundant groundwater are often encountered during construction. These bad geological bodies not only increase the difficulty of construction, but also may cause a series of serious engineering problems such as collapse, water inrush, deformation, and endanger construction safety and engineering quality. Therefore, how to accurately identify these bad geological bodies in real time during construction, especially during drilling, has become an important technical problem to be solved in the field of underground engineering.

[0003] Currently, in the process of bad geological body identification, although traditional geophysical methods such as seismic wave detection and geological radar are widely used, they have obvious limitations. First, the detection depth of geophysical methods is limited, especially in water-bearing or high-density rock layers, the signal is easily attenuated, making it difficult to accurately identify deep geological bodies. Second, the spatial resolution of these methods is insufficient, making it difficult to accurately locate bad geological bodies under complex geological conditions. In addition, geophysical data is usually obtained before construction, which cannot reflect the geological changes during construction in real time, leading to delayed decision-making.

[0004] Drilling identification technology has strong advantages in identifying bad geological bodies by collecting drilling machine operation parameters, drilling vibration parameters, and other data in real time, and has strong real-time performance, which can synchronously obtain deep underground geological information and adjust the construction scheme in time according to real-time feedback. However, the present inventors have found that the drilling identification technology relying on single modal data such as only using drilling machine parameters or acoustic signals still has the following problems: the information of single modal data is limited, which is easily affected by noise and environmental interference, leading to inaccurate identification results, especially under complex geological conditions, the adaptability is poor, and misjudgment is easy to occur. SUMMARY

[0005] In view of the technical problems in the prior art that the drilling identification technology relying on single modal data has limited information of single modal data, is easily affected by noise and environmental interference, leads to inaccurate identification results, and especially under complex geological conditions, the adaptability is poor and misjudgment is easy to occur, the present application provides a drilling bad geological body intelligent identification method based on multi-modal fusion.

[0006] In order to solve the above technical problems, the present application adopts the following technical solutions:

[0007] A method for intelligently identifying unfavorable geological bodies while drilling based on multimodal fusion includes the following steps:

[0008] Data collection: Collect multimodal while-drilling data and corresponding adverse geological body information. The multimodal while-drilling data includes drilling rig operating parameters, while-drilling vibration parameters, while-drilling acoustic signals, and while-drilling resistivity imaging.

[0009] Data preprocessing: Four modal data types, including drilling rig operating parameters, drilling vibration parameters, drilling acoustic signals, and drilling resistivity imaging, are selected to form a multimodal drilling dataset. The adverse geological body information corresponding to the drilling data is selected to form dataset labels. The dataset labels are used to guide the model in learning how to identify different adverse geological bodies based on the input multimodal drilling data. By comparing the model's predicted results with the actual geological conditions, the model can learn by comparing the predicted and actual labels, and continuously optimize and improve the model.

[0010] Feature extraction: The input of the multimodal while drilling dataset is four modes: drilling rig operating parameter mode r, while drilling vibration parameter mode v, while drilling acoustic signal mode a, while drilling resistivity imaging mode i. The corresponding original modal data is m∈{r,v,a,i}; where, T m The number of input samples for different modalities, d m are the feature dimensions of different modalities, and m is the modal category;

[0011] For the three modes of drilling rig operating parameters, drilling vibration parameters, and drilling acoustic signals, we use stacked bidirectional long short-term memory (BiLSTM) to map different original modal data into a unified dimension to obtain features. m∈{r,v,a}:x m =BiLSTM(o m ),m∈{r,v,a}; the image feature extractor residual neural network, ResNet, is used for the resistivity imaging mode while drilling to map the original modal data to a unified dimension to obtain the feature m∈{i};x m =ResNet(o m ),m∈{i}; the unified dimensional features extracted from the four modalities are x m ,m∈{r,v,a,i};

[0012] Modal common and private feature extraction: Each modal feature is mapped twice to obtain the modal common feature space and modal private feature space. Specifically, different modal features are mapped through the shared encoder E c Mapping to the common hidden features of the modalities m∈{r,v,a,i}: m∈{r,v,a,i}, the different modal features are passed through the private encoder E p Mapping to modality-private hidden features m∈{r,v,a,i}: m∈{r,v,a,i}; among them, there are a total of encoders E c and private encoder E p They are composed of feedforward neural network and Sigmoid activation function; the four modal features are respectively encoded by the common encoder E c Mapping to the modality-shared feature space and private encoder E p After mapping to the modality private feature space, the four modal common hidden features are obtained and four modal private hidden features

[0013] Modal fusion: Consider the features of the modal common subspace as the common modality Consider the features of the modality-private subspace as private modalities Stack the common mode and private mode to get the matrix According to the Cross-modal Transformer model, the shared modal features are enhanced using private modal data, and the stacking matrix M is transformed into Next, use the Transformer model to extract features and obtain the matrix

[0014] Identification of unfavorable geological bodies: Matrix vectors are concatenated to obtain a matrix After the matrix The final bad geological body recognition result is obtained through a feedforward neural network G(·) Indicates whether there are unfavorable geological bodies in the drilling area and the type of unfavorable geological bodies.

[0015] Furthermore, in the data preprocessing step, the adverse geological bodies corresponding to the drilling data include faults, caves, weak interlayers, karst, expansive soil and groundwater-rich areas.

[0016] Furthermore, in the modal fusion step, the Cross-modal Transformer model is a deep learning model for processing multimodal data. It improves the performance of the task by combining information between different modalities. Each Cross-modal Transformer consists of D layers of cross-modal attention blocks. The calculation formula of the Cross-modal Transformer is:

[0017]

[0018] Among them, fθ is a feedforward neural network parameterized by theta, represents a multi-head cross-modal attention mechanism, and LN represents layer normalization.

[0019] Further, in the adverse geological body identification step, the feedforward neural network G(·) includes sequentially connected input layer, hidden layer and output layer, and in a series of transformations of the hidden layer, each layer is provided with an activation function capable of capturing complex relationships in data to enhance the non-linear processing capability of the model.

[0020] Compared with the prior art, the multi-modal fusion based intelligent identification method for drilling adverse geological bodies provided by the present application collects drilling rig operation parameters, drilling vibration parameters, drilling acoustic signals and drilling resistivity imaging maps to form a multi-modal drilling data set, and maps the modal characteristics of the four kinds of modal data to a unified dimension using BiLSTM and ResNet; on this basis, the four kinds of modal characteristics are respectively mapped to a modal common feature space and a private encoder E c mapping to a modal common feature space and a private encoder E p mapping to a modal private feature space, to obtain four modal common hidden features and four modal private hidden features; then a Cross-modal Transformer model is used to combine private modal data with common modal characteristics to improve the performance of the task, then a Transformer model is used for feature extraction, and finally a feedforward neural network is used to realize intelligent identification of adverse geological bodies during drilling. Traditional drilling identification mainly uses single modal data, but the amount of information carried by one kind of modal data is limited and cannot contain all the information of the geological body. Multi-modal data can fully utilize the complementarity of different data sources to provide more comprehensive analysis for the identification of adverse geological bodies. This technology not only improves the accuracy and reliability of the identification, but also enhances the anti-noise ability through cross-validation of different data modalities to filter out environmental interference signals. In addition, multi-modal fusion can adapt to more complex geological environments, especially in the case of coexistence of multiple geological anomalies, it can better identify potential adverse geological bodies and provide real-time and accurate feedback. Therefore, the use of multi-modal drilling identification technology will effectively improve the safety and efficiency of underground engineering construction and provide solid technical support for construction under complex geological conditions. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a flowchart of the multi-modal fusion based intelligent identification method for drilling adverse geological bodies provided by the present application.

[0022] Figure 2 is a structure diagram of the Cross-modal Transformer model provided by the present application. DETAILED DESCRIPTION

[0023] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below with reference to specific illustrations.

[0024] Please refer to Figure 1 and Figure 2 As shown, the present invention provides a method for intelligently identifying unfavorable geological bodies while drilling based on multimodal fusion, comprising the following steps:

[0025] Data collection: Collect multimodal downhole data and corresponding adverse geological body information. The multimodal downhole data includes drilling rig operating parameters, downhole vibration parameters, downhole acoustic signals and downhole resistivity imaging. Specifically, the drilling rig operating parameters collect the drilling rig's working status data in real time through built-in sensors or external monitoring equipment, including speed, thrust, torque, drilling speed, etc.; the downhole vibration parameters monitor the vibration of the drill pipe in real time by installing an acceleration sensor near the drill bit; the downhole acoustic signal collects the acoustic signal generated by the drill bit breaking the rock in real time by installing an acoustic sensor near the drill bit; the downhole resistivity imaging is obtained in real time by installing a downhole resistivity imaging logging instrument near the drill bit.

[0026] Data preprocessing: Four modal data types, namely, drilling rig operating parameters, while-drilling vibration parameters, while-drilling acoustic signals, and while-drilling resistivity images, are selected to form a multimodal while-drilling dataset. The adverse geological body information corresponding to the while-drilling data is selected to form dataset labels. The dataset labels are used to guide the model in learning how to identify different adverse geological bodies based on the input multimodal while-drilling data. By comparing the model's predicted results with the actual geological conditions, the model can learn by comparing the predicted and actual labels, and continuously optimize and improve the model. Specifically, adverse geological bodies include faults, caves, weak interlayers, karst, expansive soil, and areas with abundant groundwater.

[0027] Feature extraction: The input of the multimodal while drilling dataset is four modes: drilling rig operating parameter mode r, while drilling vibration parameter mode v, while drilling acoustic signal mode a, while drilling resistivity imaging mode i. The corresponding original modal data is m∈{r,v,a,i}; where, T m The number of input samples for different modalities, d m are the feature dimensions of different modalities, and m is the modal category;

[0028] For the drilling rig operating parameters, while drilling vibration parameters, and while drilling acoustic signal modes, these three modes belong to low-dimensional modes, and the while drilling data has time series, so the stacked bidirectional long short-term memory (BiLSTM) is used to map different original modal data to a unified dimension to obtain features. m∈{r,v,a}:x m =BiLSTM(om ), m e {r, v, a}; for the resistivity imaging while drilling modal, use the image feature extractor residual neural network, ResNet, to map the original modal data to a unified dimension to obtain the feature m e {i}; x m = ResNet(o m ), m e {i}; the final four modal extracted unified dimension features are x m , m e {r, v, a, i}.

[0029] Modal common and private feature extraction: the geological body characteristics while drilling are reflected in drilling rig operating parameters, while drilling vibration parameters, while drilling acoustic signals, and while drilling resistivity imaging, each modal information contains information useful for judging the geological body characteristics, which belongs to the information common to each modal; at the same time, the data structure and data form of each modal data have great differences, so the information of different modal contains the information private to each modal. The features extracted by drilling rig operating parameters, while drilling vibration parameters, while drilling acoustic signals, and while drilling resistivity imaging are x m , m e {r, v, a, i}, each modal feature is mapped twice to obtain the modal common feature space and the modal private feature space, specifically: different modal features are mapped to the modal common hidden feature c by the common encoder E m e {r, v, a, i}: m e {r, v, a, i}, the common encoder E c is composed of a feedforward neural network and a Sigmoid activation function, the common hidden feature is considered to be the information useful for judging the geological body characteristics in different modal features, because this part of information has the same tendency in the analysis of the geological body characteristics, and in the common feature space, it shows a similar distance, and the common encoders of different modal use the same parameters

[0030] Different modal features are mapped to the modal private hidden feature p by the private encoder E m e {r, v, a, i}: m e {r, v, a, i}, the private encoder E p is also composed of a feedforward neural network and a Sigmoid activation function, the private hidden feature is considered to be the information private to each modal in judging the geological body characteristics, because the private features of each modal have great differences, and have a far distance in the private feature space, so the private encoders of different modal use different parameters

[0031] The four modal features are respectively encoded by the common encoder E c Mapping to the modality-shared feature space and private encoder E p After mapping to the modality private feature space, the four modal common hidden features are obtained and four modal private hidden features

[0032] Modal fusion: Consider the features of the modal common subspace as the common modality Consider the features of the modality-private subspace as private modalities Stack the common mode and private mode to get the matrix According to the Cross-modal Transformer model, the shared modal features are enhanced using private modal data, and the stacking matrix M is transformed into

[0033] The Cross-modal Transformer model is a deep learning model for processing multimodal data. It improves the performance of tasks by combining information between different modalities. Each Cross-modal Transformer consists of D layers of cross-modal attention blocks. The calculation formula of the Cross-modal Transformer is:

[0034]

[0035] Among them, f θ is a feedforward neural network parameterized by θ, represents the multi-head cross-modal attention mechanism, and LN represents layer normalization;

[0036] Next, use the Transformer model to extract features and obtain the matrix

[0037] Identification of unfavorable geological bodies: Matrix vectors are concatenated to obtain a matrix After the matrix The final bad geological body recognition result is obtained through a feedforward neural network G(·) Indicates whether there are bad geological bodies in the drilling area and the type of bad geological bodies. The feedforward neural network G(·) transmits and processes information through a series of hierarchical neurons. In the input layer, the original feature data is input into the network, and then these data are transformed through a series of hidden layers. Each layer uses an activation function to enhance the nonlinear processing ability of the model, so that the complex relationship in the data can be captured. Finally, the bad geological body identification result is obtained at the output layer. That is, the feedforward neural network G(·) includes sequentially connected input layers, hidden layers, and output layers, and each layer in a series of transformations of the hidden layers is provided with an activation function that enhances the nonlinear processing capability of the model and can capture complex relationships in the data.

[0038] Compared with the prior art, the multi-modal fusion-based intelligent identification method for drilling adverse geological bodies provided by the present application collects drilling operation parameters, drilling vibration parameters, drilling acoustic signals, and drilling resistivity imaging maps to form a multi-modal drilling data set, maps the modal characteristics of the four modal data to a unified dimension using BiLSTM and ResNet, and on this basis, maps the four modal characteristics to a modal common feature space and a modal private feature space through a common encoder E c and a private encoder E p respectively, to obtain four modal common hidden characteristics and four modal private hidden characteristics; then uses a Cross-modal Transformer model to combine private modal data with common modal characteristics to improve the performance of the task, then uses a Transformer model for feature extraction, and finally realizes intelligent identification of adverse geological bodies while drilling through a feedforward neural network. Traditional drilling identification mainly uses single modal data, but the amount of information carried by one kind of modal data is limited and cannot contain all the information of the geological body. Multi-modal data can fully utilize the complementarity of different data sources to provide more comprehensive analysis for the identification of adverse geological bodies. This technology not only improves the accuracy and reliability of the identification, but also enhances the anti-noise ability through cross-validation of different data modalities to filter out environmental interference signals. In addition, multi-modal fusion can adapt to more complex geological environments, especially in the case of multiple geological anomalies coexisting, it can better identify potential adverse geological bodies and provide real-time and accurate feedback. Therefore, the use of multi-modal drilling identification technology will effectively improve the safety and efficiency of underground engineering construction and provide solid technical support for construction under complex geological conditions.

[0039] Finally, it should be pointed out that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the purpose and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A method for intelligent identification of unfavorable geological bodies while drilling based on multimodal fusion, characterized in that: The following steps are involved: Data collection: Collect multimodal while-drilling data and corresponding adverse geological body information. The multimodal while-drilling data includes drilling rig operating parameters, while-drilling vibration parameters, while-drilling acoustic signals, and while-drilling resistivity imaging. Data preprocessing: Four modal data types, including drilling rig operating parameters, drilling vibration parameters, drilling acoustic signals, and drilling resistivity imaging, are selected to form a multimodal drilling dataset. The unfavorable geological bodies corresponding to the drilling data are selected to form dataset labels. The dataset labels are used to guide the model in learning how to identify different unfavorable geological bodies based on the input multimodal drilling data. By comparing the model's predicted results with the actual geological conditions, the model can learn by comparing the predicted and actual labels, and continuously optimize and improve the model. Feature extraction: The input of the multimodal while drilling dataset is four modes: drilling rig operating parameter mode r, while drilling vibration parameter mode v, while drilling acoustic signal mode a, while drilling resistivity imaging mode i. The corresponding original modal data is Among them, T m The number of input samples for different modalities, d m are the feature dimensions of different modalities, and m is the modal category; For the three modes of drilling rig operating parameters, drilling vibration parameters, and drilling acoustic signals, we use stacked bidirectional long short-term memory (BiLSTM) to map different original modal data into a unified dimension to obtain features. x m =BiLSTM(o m ),m∈{r,v,a}; the image feature extractor residual neural network, ResNet, is used for the resistivity imaging mode while drilling to map the original modal data to a unified dimension to obtain the feature x m =ResNet(o m ),m∈{i}; the unified dimensional features extracted from the four modalities are x m ,m∈{r,v,a,i}; Modal common and private feature extraction: Each modal feature is mapped twice to obtain the modal common feature space and modal private feature space. Specifically, different modal features are mapped through the shared encoder E c Mapping to the common hidden features of the modalities The different modal features are passed through the private encoder E p Mapping to modality-private hidden features Among them, there are a total of encoders E c and private encoder E p They are composed of feedforward neural network and Sigmoid activation function; the four modal features are respectively encoded by the common encoder E c Mapping to the modality-shared feature space and private encoder E p After mapping to the modality private feature space, the four modal common hidden features are obtained and four modal private hidden features Modal fusion: Consider the features of the modal common subspace as the common modality Consider the features of the modality-private subspace as private modalities Stack the common mode and private mode to get the matrix According to the Cross-modal Transformer model, the shared modal features are enhanced using private modal data, and the stacking matrix M is transformed into Next, use the Transformer model to extract features and obtain the matrix Identification of unfavorable geological bodies: Matrix vectors are concatenated to obtain a matrix After the matrix The final bad geological body recognition result is obtained through a feedforward neural network G(·) Indicates whether there is any adverse geological body information in the drilling area and the type of adverse geological body.

2. The method for intelligent identification of unfavorable geological bodies while drilling based on multimodal fusion according to claim 1, characterized in that: In the data preprocessing step, the adverse geological bodies corresponding to the drilling data include faults, caves, weak interlayers, karst, expansive soil and groundwater-rich areas.

3. The method for intelligent identification of unfavorable geological bodies while drilling based on multimodal fusion according to claim 1, characterized in that: In the modal fusion step, the Cross-modal Transformer model is a deep learning model for processing multimodal data. It improves the performance of the task by combining information between different modalities. Each Cross-modal Transformer consists of D layers of cross-modal attention blocks. The calculation formula of the Cross-modal Transformer is: Among them, f θ is a feedforward neural network parameterized by θ, represents the multi-head cross-modal attention mechanism, and LN represents layer normalization.

4. The method for intelligent identification of unfavorable geological bodies while drilling based on multimodal fusion according to claim 1, characterized in that: In the step of identifying adverse geological bodies, the feedforward neural network G(·) includes a sequentially connected input layer, a hidden layer, and an output layer, and in a series of transformations of the hidden layer, each layer is provided with an activation function that enhances the nonlinear processing capability of the model, thereby being able to capture complex relationships in the data.

Images