A method for identifying a through-hole breakpoint based on multi-source data and a convolutional neural network

Abstract

The application discloses a kind of based on multi-source data and convolutional neural network's crossing well breakpoint identification method, comprising: obtaining target well drilling parameter, well logging curve and recording well parameter, standardization pretreatment is carried out;Based on preprocessed data, by constructing derivative parameter and multi-scale sliding window statistical characteristics, generate high-dimensional feature dataset;Based on well logging-seismic cross-validation result, for feature data label fault class label, construct training sample library;With sample library training combination bayesian hyperparameter optimization's convolutional neural network model, obtain intelligent identification model;The data of well section to be identified are input into the model after the same process, and the steps such as breakpoint position and fault type probability are output.The present application comprehensively utilizes drilling-well logging-recording well multi-source high-resolution data, realizes the automatic, high-precision identification of the micro fault crossing well that is difficult to identify by seismic, and provides technical support for the fine development of complex fault block reservoir.

Description

Technical Field

[0001] This invention belongs to the field of geophysical logging and downhole geological structure identification technology in oil and gas exploration and development, specifically involving a method for identifying well breakpoints based on multi-source data and convolutional neural networks. Background Technology

[0002] In the development of complex fault-block oilfields, such as the Beibu Gulf Basin in the western South China Sea, accurately characterizing micro-faults is crucial for optimizing horizontal well trajectories, improving reservoir encounter rates, mitigating drilling risks, and identifying geological reserves. Currently, downhole fault identification primarily relies on 3D seismic data, interpreting it using seismic attributes such as coherence, curvature, and ant-body. However, the vertical resolution of seismic data is limited, often making it difficult to clearly identify small faults or fault points with displacements smaller than the seismic tuning thickness (typically 10-20 meters). This introduces uncertainty into the precise implementation of development wells.

[0003] On the other hand, drilling, logging, and well logging (hereinafter referred to as "drilling-logging-well logging") data originate directly from the wellbore and have a much higher vertical resolution than seismic data. When the wellbore passes through a fault, it can cause various abnormal response characteristics in the data, such as: abrupt changes or distortions in logging curves (e.g., density, resistivity); abnormal fluctuations in drilling parameters (e.g., drilling pressure, torque, rotational speed); and duplication or omissions in logging lithology. Traditionally, these characteristics have been identified by manual comparative analysis, but this method is inefficient, highly subjective, and ineffective in identifying closed faults with weak characteristics.

[0004] In recent years, machine learning technology has been applied in the field of geophysics. However, there is still a lack of publicly reported systematic methods that can comprehensively utilize multi-source data from drilling, logging, and surveying to achieve intelligent identification of well faults through convolutional neural networks (CNNs). Existing technologies mostly rely on single-type data and fail to effectively integrate multi-scale features, and even more so lack automated, high-precision identification schemes for small fault samples. Therefore, developing a well fault identification method that can integrate multi-source high-resolution data, automatically extract fault-sensitive features, and achieve intelligent identification has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies, such as insufficient accuracy in identifying micro-faults due to reliance on seismic data, and low efficiency and strong subjectivity of traditional well logging comparison methods. Its purpose is to provide a well-crossing fault identification method based on multi-source data and convolutional neural networks.

[0006] It should be noted that in this invention, "through-well fault" refers to a fault structure through which the wellbore passes, while "through-well breakpoint" refers to the specific depth point where the fault intersects with the wellbore. The two are essentially the same; the "through-well breakpoint identification" described in this invention is the precise determination of the location of the through-well fault.

[0007] This invention is achieved through the following technical solution: A method for identifying well breakpoints based on multi-source data and convolutional neural networks includes the following steps: S1. Multi-source data acquisition and preprocessing: Acquire drilling parameters, logging curves and logging parameters of the target well section to form the original multi-source dataset, and preprocess the original multi-source dataset; The drilling and logging parameters are obtained from the drilling database and include bit weight (WOB), rate of drilling (ROP), torque (TQA), rotational speed (RPM), pump pressure (SPPA), pump displacement (MFIA), mud return rate (MFOA), etc.

[0008] The logging curves are extracted from the logging interpretation results and include natural gamma (GR), shallow and deep lateral resistivity (RD, RS), bulk density (DEN), neutron porosity (CNL), acoustic transit time (AC), photoelectric absorption cross section index (PE), etc.

[0009] The preprocessing includes depth alignment, outlier removal, and normalization. The depth alignment ensures that various types of data correspond accurately at the same depth point; The outlier removal refers to removing obvious outliers caused by instrument malfunctions, etc. The standardization process employs Z-score standardization to process all parameters of the original multi-source dataset, eliminating the influence of dimensions and ensuring all features are on the same order of magnitude. The formula for Z-score standardization is: In the formula: μ is the mean of the feature, and σ is the standard deviation of the feature; S2. Construction of Multi-Scale Fault Sensitive Features: Step S2 is the core of improving the sensitivity of small fault identification. Anomalies in a single parameter may be caused by faults, lithological changes or other geological factors, and there are multiple solutions. This invention constructs derived features by fusing multiple parameters to amplify the fault response signal and improve the accuracy of identification.

[0010] Based on the preprocessed multi-source dataset, derived feature parameters that can amplify tomographic signals and reduce the ambiguity of identification are constructed through multi-parameter fusion. Furthermore, multi-scale sliding window statistics are performed on the preprocessed multi-source data and derived feature parameters along the depth direction to generate a high-dimensional feature dataset. The derived characteristic parameters include at least one of mechanical specific energy, drilling efficiency parameters, torque-to-speed ratio, mud return ratio, rock drillability index, density-neutron product, and depth-to-shallow resistivity ratio.

[0011] The mechanical specific energy comprehensively reflects the energy required to break a unit volume of rock. It typically decreases in fault fracture zones. The formula for calculating mechanical specific energy is: In the formula: MSE is mechanical specific energy, in megapascals (MPa); WOB is drilling pressure, in kilonewtons (kN); RPM is rotational speed, in revolutions per minute (r / min); ROP is drilling speed, in meters per hour (m / h); TQA is torque, in kilonewton-meters (kN·m); D is wellbore diameter, in inches (in). The drilling efficiency parameter is the ratio of drilling pressure to drilling speed. This parameter directly reflects the drillability of rocks; rocks in fault fracture zones are softer and have better drillability. The torque-to-speed ratio is the ratio of torque to speed. This parameter reflects the frictional resistance encountered by the drill bit when cutting rock. Loose rock structures or the presence of fillings with different friction coefficients within the fault zone can cause abnormal fluctuations in this ratio. The mud return ratio is the ratio of mud return volume to pump discharge volume. This parameter is an important indirect indicator for fault identification; if the fault fracture zone is connected to the underground high-pressure fluid layer or seepage layer, it will directly lead to abnormal return volume (well kick or well leakage). The rock drillability index is the mechanical specific energy. Divide by mud content and porosity sum ( This parameter is a composite index of rock drillability. By dividing by "mud content + porosity", the influence of lithology and primary porosity on mechanical specific energy is eliminated, thereby improving the accuracy of identification. The density-neutron product is the product of volume density and neutron porosity. This parameter comprehensively reflects lithology and porosity. Faulting may cause rock fracturing or produce specific infill materials, and this product can sensitively reflect this combined effect. The deep-to-shallow resistivity ratio is the ratio of the deep lateral resistivity to the shallow lateral resistivity. This parameter is a key indicator for identifying permeable layers and judging fluid properties; fluid activity within fault fracture zones may cause systematic changes in the difference between the resistivity of the original formation and the resistivity of the intrusive zone.

[0012] The specific method for the multi-scale sliding window statistics is as follows: S21. Determine the optimal preset window length L: Multiple preset window lengths are set, and the preprocessed multi-source data and derived feature parameters are used as inputs. High-dimensional feature datasets are constructed using multiple different window lengths, and a preliminary convolutional neural network model is trained. The F1 score of the convolutional neural network model for the tomography identification task is used as the evaluation index, and the window length with the highest F1 score is selected as the optimal preset window length L for the final application. The F1 score is a commonly used metric in machine learning for comprehensively evaluating the accuracy of classification models. It is the harmonic mean of precision and recall, and its calculation formula is as follows: In this invention, the F1 score is used to comprehensively balance the accuracy and coverage of tomography identification, which is especially suitable for small sample classification and identification tasks such as tomography identification.

[0013] S22. Using the optimal preset window length L, for each parametric curve, slide the window point by point along the depth with a fixed step size to calculate the high-dimensional feature dataset within each window.

[0014] The high-dimensional feature dataset includes the mean, standard deviation, slope, skewness, and kurtosis statistics. The mean reflects the average level of the parameters within the window and can be used to identify systematic shifts in lithology and physical properties caused by fault zones. The standard deviation reflects the oscillation amplitude of the parameters around the mean; abrupt changes in the physical properties of rocks in fault zones can lead to an increase in the standard deviation. The slope reflects the trend of parameter changes with depth, and abrupt changes may occur at fault interfaces. Skewness reflects the asymmetry of parameter distribution and can reveal the distortion of parameter distribution caused by fault activity. Kurtosis reflects the sharpness of the parameter distribution pattern and is related to the suddenness of faulting.

[0015] S3. Construct a fault sample library combining well logging and seismic analysis: Based on the fault information interpreted from the seismic profile and the high-dimensional feature dataset generated in step S2, cross-validation of well logging and seismic analysis at the same depth segment is used to label the high-dimensional feature dataset samples with category labels, forming a labeled training sample library. The purpose of step S3 is to provide high-quality labels for supervised learning. In this invention, to construct training samples for supervised learning, the sample points are divided into three categories of labels based on the degree of agreement between seismic interpretation results and downhole multi-parameter responses. These three categories of labels are determined by geological and well logging interpretation experts, aiming to provide the model with high-quality and reliable fault identification and classification data.

[0016] The category labels include Class I fault labels, Class II fault labels, and Class III fault labels; Type I fault labels indicate that both seismic and downhole data show obvious fault responses, that is, there is obvious dislocation or distortion of the same phase axis on the seismic profile, and there are obvious abnormal responses in the drilling-logging-well logging multi-source features at the corresponding depth (the same phase axis on the seismic profile shows clear discontinuity, and multiple drilling, logging and well logging parameters at the corresponding depth show synchronous anomalies that conform to the fault response law). Type II fault labels indicate weak fault responses in both seismic and downhole data. This means that there are suspicious minor fault movements or weakened phase axes in the seismic profile, or weak responses in the drill-scan-logging features (suspicious faults or local disturbances in the phase axes in the seismic profile, or abnormal indications in multiple downhole parameters, but the sufficiency of the evidence is weaker than that of Type I samples). Class III fault label indicates that there is no obvious fault response in both seismic and downhole data, that is, there are no obvious anomalies in the seismic profile and drilling-logging multi-source characteristics (the seismic profile has good continuity and the downhole multi-parameter curves are stable without abnormal fluctuations). S4. Training a Bayesian-optimized convolutional neural network intelligent identification model for well breaks: Construct a convolutional neural network model and use the Bayesian optimization algorithm to optimize the hyperparameter space of the convolutional neural network model to obtain the optimal model; use the labeled training sample library obtained in step S3 to train the optimal model to obtain the intelligent identification model for well breaks. The convolutional neural network adopts a one-dimensional convolutional structure to extract feature patterns along the depth direction; the hyperparameters sought by Bayesian optimization include the number of network layers, the number of filters in each convolutional layer, the number of neurons in the fully connected layer, and the learning rate of the model training.

[0017] S5. Apply the intelligent identification model for fault types using the well-crossing fault point identification model: After processing the data of the well section to be identified through steps S1 and S2, input it into the intelligent identification model for well-crossing fault points trained in step S4. The intelligent identification model for well-crossing fault points outputs the fault identification results along the well depth and the probability distribution of each depth point belonging to different fault categories.

[0018] The beneficial effects of this invention are: This invention provides a well-through fault identification method based on multi-source data and convolutional neural networks. By fusing high-resolution data from drilling, logging, and well logging sources, a multi-scale feature engineering model that can amplify fault response signals is constructed. Combined with a Bayesian-optimized convolutional neural network model, this method achieves high-precision and automated identification of small faults (breakpoints) passing through wells. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of multi-scale sliding window feature extraction in Embodiment 1 of the present invention; Figure 3This is a schematic diagram illustrating the application effect of Embodiment 1 of the present invention.

[0020] For those skilled in the art, other related figures can be obtained from the above figures without any creative effort. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1

[0023] This embodiment uses the identification of well-crossing fault points in a complex fault-block oilfield in the Beibu Gulf Basin of the South China Sea as an example for detailed explanation.

[0024] like Figure 1 As shown, a method for identifying well-crossing breakpoints based on multi-source data and convolutional neural networks includes the following steps: S1. Multi-source data acquisition and preprocessing: Acquire drilling parameters, logging curves and logging parameters of the target well section to form the original multi-source dataset, and preprocess the original multi-source dataset; The drilling and logging parameters are obtained from the drilling database and include bit weight (WOB), rate of drilling (ROP), torque (TQA), rotational speed (RPM), pump pressure (SPPA), pump displacement (MFIA), mud return rate (MFOA), etc.

[0025] The logging curves are extracted from the logging interpretation results and include natural gamma (GR), shallow and deep lateral resistivity (RD, RS), bulk density (DEN), neutron porosity (CNL), acoustic transit time (AC), photoelectric absorption cross section index (PE), etc.

[0026] The preprocessing includes depth alignment, outlier removal, and normalization. The depth alignment ensures that various types of data correspond accurately at the same depth point; The outlier removal refers to removing obvious outliers caused by instrument malfunctions, etc. The standardization process employs Z-score standardization to process all parameters of the original multi-source dataset, eliminating the influence of dimensions and ensuring all features are on the same order of magnitude. The formula for Z-score standardization is: In the formula: μ is the mean of the feature, and σ is the standard deviation of the feature; S2. Construction of Multi-Scale Fault Sensitive Features: Step S2 is the core of improving the sensitivity of small fault identification. Anomalies in a single parameter may be caused by faults, lithological changes or other geological factors, and there are multiple solutions. This invention constructs derived features by fusing multiple parameters to amplify the fault response signal and improve the accuracy of identification.

[0027] Based on the preprocessed multi-source dataset, derived feature parameters that can amplify tomographic signals and reduce the ambiguity of identification are constructed through multi-parameter fusion. Furthermore, multi-scale sliding window statistics are performed on the preprocessed multi-source data and derived feature parameters along the depth direction to generate a high-dimensional feature dataset. The derived characteristic parameters include at least one of mechanical specific energy, drilling efficiency parameters, torque-to-speed ratio, mud return ratio, rock drillability index, density-neutron product, and depth-to-shallow resistivity ratio.

[0028] The mechanical specific energy comprehensively reflects the energy required to break a unit volume of rock. It typically decreases in fault fracture zones. The formula for calculating mechanical specific energy is: In the formula: MSE is mechanical specific energy, in megapascals (MPa); WOB is drilling pressure, in kilonewtons (kN); RPM is rotational speed, in revolutions per minute (r / min); ROP is drilling speed, in meters per hour (m / h); TQA is torque, in kilonewton-meters (kN·m); D is wellbore diameter, in inches (in). The drilling efficiency parameter is the ratio of drilling pressure to drilling speed. This parameter directly reflects the drillability of rocks; rocks in fault fracture zones are softer and have better drillability. The torque-to-speed ratio is the ratio of torque to speed. This parameter reflects the frictional resistance encountered by the drill bit when cutting rock. Loose rock structures or the presence of fillings with different friction coefficients within the fault zone can cause abnormal fluctuations in this ratio. The mud return ratio is the ratio of mud return volume to pump discharge volume. This parameter is an important indirect indicator for fault identification; if the fault fracture zone is connected to the underground high-pressure fluid layer or seepage layer, it will directly lead to abnormal return volume (well kick or well leakage). The rock drillability index is the mechanical specific energy. Divide by mud content and porosity sum ( This parameter is a composite index of rock drillability. By dividing by "mud content + porosity", the influence of lithology and primary porosity on mechanical specific energy is eliminated, thereby improving the accuracy of identification. The density-neutron product is the product of volume density and neutron porosity. This parameter comprehensively reflects lithology and porosity. Faulting may cause rock fracturing or produce specific infill materials, and this product can sensitively reflect this combined effect. The deep-to-shallow resistivity ratio is the ratio of the deep lateral resistivity to the shallow lateral resistivity. This parameter is a key indicator for identifying permeable layers and judging fluid properties; fluid activity within fault fracture zones may cause systematic changes in the difference between the resistivity of the original formation and the resistivity of the intrusive zone.

[0029] Sliding window statistics were performed with a given window length. Using the original and derived parameters as input, the recognition performance (indicated by F1 score) of the preliminary model was compared under different preset window lengths (e.g., 5m, 10m, 15m, 20m, 25m). The optimal sliding window length (L) for this study area was determined to be 20 meters. Figure 2 As shown, following a given optimal window length, the system slides point by point along the depth range of the parameters. Within each window, five statistics are calculated for each parameter curve: Mean: The average level of parameters within the window, which can be used to identify systematic shifts in lithology and physical properties caused by fault zones.

[0030] Standard deviation: The amplitude of a parameter's oscillation around its mean. Abrupt changes in the physical properties of rocks in fault zones can lead to an increase in standard deviation.

[0031] Slope: Reflects the trend of parameter change with depth, and may change abruptly at the fault interface.

[0032] Skewness: reflects the asymmetry of parameter distribution and can reveal the morphological distortion of parameter distribution caused by fault activity.

[0033] Kurtosis: Reflects the sharpness of the parameter distribution pattern and is related to the suddenness of faulting.

[0034] The F1 score is a commonly used metric in machine learning for comprehensively evaluating the accuracy of classification models. It is the harmonic mean of precision and recall, and its calculation formula is as follows: The F1 score is used to comprehensively balance the accuracy and coverage of tomography, making it particularly suitable for small-sample classification tasks such as tomography.

[0035] Thus, each depth point is transformed into an M×5 dimensional feature vector, where M is the total number of input parameter curves (M=21 in this embodiment), resulting in a 105-dimensional feature vector, which greatly enriches the data representation capabilities.

[0036] S3. Construct a fault sample library combining well logging and seismic analysis: Based on the fault information interpreted from the seismic profile and the high-dimensional feature dataset generated in step S2, cross-validation of well logging and seismic analysis at the same depth segment is used to label the high-dimensional feature dataset samples with category labels, forming a labeled training sample library. The purpose of step S3 is to provide high-quality labels for supervised learning. In this invention, to construct training samples for supervised learning, the sample points are divided into three categories of labels based on the degree of agreement between seismic interpretation results and downhole multi-parameter responses. These three categories of labels are determined by geological and well logging interpretation experts, aiming to provide the model with high-quality and reliable fault identification and classification data.

[0037] The category labels include Class I fault labels, Class II fault labels, and Class III fault labels; Type I fault labels indicate that both seismic and downhole data show obvious fault responses, that is, there is obvious dislocation or distortion of the same phase axis on the seismic profile, and there are obvious abnormal responses in the drilling-logging-well logging multi-source features at the corresponding depth (the same phase axis on the seismic profile shows clear discontinuity, and multiple drilling, logging and well logging parameters at the corresponding depth show synchronous anomalies that conform to the fault response law). Type II fault labels indicate weak fault responses in both seismic and downhole data. This means that there are suspicious minor fault movements or weakened phase axes in the seismic profile, or weak responses in the drill-scan-logging features (suspicious faults or local disturbances in the phase axes in the seismic profile, or abnormal indications in multiple downhole parameters, but the sufficiency of the evidence is weaker than that of Type I samples). Class III fault label indicates that there is no obvious fault response in both seismic and downhole data, that is, there are no obvious anomalies in the seismic profile and drilling-logging multi-source characteristics (the seismic profile has good continuity and the downhole multi-parameter curves are stable without abnormal fluctuations). By meticulously classifying data from more than 10 wells in the study area, a high-quality fault identification training sample library containing depths exceeding 2000 meters was constructed.

[0038] S4. Training a Bayesian-optimized convolutional neural network intelligent identification model for well breaks: Construct a convolutional neural network model and use the Bayesian optimization algorithm to optimize the hyperparameter space of the convolutional neural network model to obtain the optimal model; use the labeled training sample library obtained in step S3 to train the optimal model to obtain the intelligent identification model for well breaks. To achieve optimal model performance, a Bayesian optimization algorithm is used to automatically optimize the hyperparameters of the CNN model. The hyperparameters include the number of convolutional layers, the number of filters in each layer, the number of neurons in the fully connected layer, and the learning rate. A reasonable search space is set for each parameter (e.g., the number of convolutional layers [2,4], the number of filters [32,128], etc.). The optimization process aims at the recognition accuracy of the model on the independent validation set, and the optimal combination of hyperparameters is determined through iterative search. The final CNN prediction model is constructed using this optimal combination. Its structure includes: an input layer, several one-dimensional convolutional and pooling layers (for extracting local depth patterns), dropout layers (to prevent overfitting), fully connected layers (for comprehensive judgment), and a Softmax output layer (outputting the probabilities of the three types of tomographic labels). The sample library constructed in step S3 is divided into training, validation, and test sets in a ratio of 70%:15%:15%. The model is fully trained using the training set, and the performance is monitored on the validation set. After training, the model achieves an F1 score of 85.7% for Class I tomographic recognition on the test set, demonstrating its effectiveness.

[0039] S5. Apply the intelligent fault identification model to identify fault types: For a new well to be evaluated, repeat steps S1 and S2 to obtain its standardized high-dimensional feature data. Input this data into the intelligent identification model trained and fixed with parameters in S4. The model performs forward calculations for each depth point and outputs a three-dimensional vector (e.g., [0.84, 0.12, 0.04]) representing the probability that the point belongs to one of the three categories: "Type I fault", "Type II fault", or "Type III fault".

[0040] The category with the highest probability is determined as the fault type at that depth, ultimately forming a continuous fault type curve along the well depth. For example... Figure 3 As shown, the second line from the right displays the probability curves of the three types of faults continuously output by the model along the well depth, while the first line from the right is the continuous fault type curve generated according to the maximum probability determination rule. This area is filled with different colored blocks: red blocks represent the determined "Type I faults", green blocks represent "Type II faults", and blue blocks represent "Type III faults".

[0041] The method of this invention enables automated and continuous identification of faults passing through wells.

[0042] In this embodiment, the method of the present invention successfully identified hidden faults not clearly indicated in the seismic profile, providing a direct basis for the refined geological understanding and development adjustment of the oil field.

[0043] The applicant declares that the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention fall within the protection scope of the present invention.

Claims

1. A method for identifying well breakpoints based on multi-source data and convolutional neural networks, characterized in that: Includes the following steps: S1. Obtain drilling parameters, logging curves and logging parameters of the target well section to form the original multi-source dataset, and preprocess the original multi-source dataset; S2. Based on the multi-source dataset preprocessed in step S1, construct derived feature parameters, and perform multi-scale sliding window statistics on the preprocessed multi-source data and derived feature parameters along the depth direction to generate a high-dimensional feature dataset. S3. Based on the fault information interpreted from the seismic profile and the high-dimensional feature dataset generated in step S2, the high-dimensional feature dataset samples are labeled with category labels through well logging-seismic cross-validation at the same depth segment, forming a labeled training sample library. S4. Construct a convolutional neural network model, use the Bayesian optimization algorithm to optimize the hyperparameters of the convolutional neural network model, and use the labeled training sample library obtained in step S3 to train the model to obtain the intelligent identification model of well breakpoints. S5. Use the intelligent identification model of well fault points obtained in step S4 to intelligently identify fault types.

2. The well-crossing breakpoint identification method based on multi-source data and convolutional neural networks according to claim 1, characterized in that: The preprocessing in step S1 includes depth alignment, outlier removal, and standardization.

3. The well-crossing breakpoint identification method based on multi-source data and convolutional neural networks according to claim 1, characterized in that: The derived characteristic parameters include at least one of mechanical specific energy, drilling efficiency parameters, torque-to-speed ratio, mud return ratio, rock drillability index, density-neutron product, and depth-to-shallow resistivity ratio.

4. The well-crossing breakpoint identification method based on multi-source data and convolutional neural networks according to claim 1, characterized in that: The specific method for the multi-scale sliding window statistics is as follows: S21. Set multiple different preset window lengths, take the preprocessed multi-source data and derived feature parameters as input, construct high-dimensional feature datasets using multiple different window lengths respectively, and train a preliminary convolutional neural network model. Use the F1 score of the convolutional neural network model for the tomography identification task as the evaluation index, and select the window length with the highest F1 score as the optimal preset window length L for the final application. The formula for calculating the F1 score is: In the formula: Precision is the accuracy rate; Recall is the recall rate; S22. Using the optimal preset window length L, for each parametric curve, slide the window point by point along the depth with a fixed step size to calculate the high-dimensional feature dataset within each window.

5. The well-crossing breakpoint identification method based on multi-source data and convolutional neural networks according to claim 1, characterized in that: The high-dimensional feature dataset includes the mean, standard deviation, slope, skewness, and kurtosis statistics.

6. The well-crossing breakpoint identification method based on multi-source data and convolutional neural networks according to claim 1, characterized in that: The category labels include Class I fault labels, Class II fault labels, and Class III fault labels; Class I fault labels indicate that both seismic and downhole data show obvious fault responses; Class II fault labels indicate that both seismic and downhole data show weak fault responses; and Class III fault labels indicate that neither seismic nor downhole data show obvious fault responses.

7. The well-crossing breakpoint identification method based on multi-source data and convolutional neural networks according to claim 1, characterized in that: Step S4 specifically involves: constructing a convolutional neural network model and using a Bayesian optimization algorithm to optimize the hyperparameter space of the convolutional neural network model to obtain the optimal model; using the labeled training sample library obtained in step S3 to train the optimal model to obtain the intelligent identification model for well breakpoints.

8. The well-crossing breakpoint identification method based on multi-source data and convolutional neural networks according to claim 7, characterized in that: The convolutional neural network adopts a one-dimensional convolutional structure to extract feature patterns along the depth direction; the hyperparameters sought by Bayesian optimization include the number of network layers, the number of filters in each convolutional layer, the number of neurons in the fully connected layer, and the learning rate of the model training.

9. The well-crossing breakpoint identification method based on multi-source data and convolutional neural networks according to claim 1, characterized in that: Specifically, step S5 involves processing the data of the well section to be identified through steps S1 and S2, and then inputting it into the intelligent identification model for well faults trained in step S4. The intelligent identification model for well faults outputs the identification results of faults along the well depth and the probability distribution of each depth point belonging to different fault categories.

Images