Method and apparatus for determining drilling risks

Abstract

The present specification provides a method and device for determining drilling risks. Before implementation, a graph autoencoder is introduced and used to train a preset detection model based on physical constraints and good application effects through unsupervised learning. During implementation, target well logging data of a target well at the present time and a well logging data set of the present time period can be obtained first; and the current working condition is determined according to the target well logging data; then the target dynamic threshold for the current target well based on the working condition constraint and considering the change of the processed data in the current time period before the graph autoencoding is determined according to the current working condition, the well logging data set of the current time period, and a preset geology-engineering pre-drilling evaluation profile; the corresponding target reconstruction error is obtained by processing the target well logging data using the preset detection model; and then whether there is a drilling risk is detected according to the target reconstruction error and the target dynamic threshold. Thus, the drilling risk can be accurately detected and identified, and the false positive rate is reduced.

Description

Technical Field

[0001] This manual pertains to the field of oil and gas drilling technology, and particularly relates to methods and apparatus for determining drilling risks. Background Technology

[0002] As shallow oil and gas resources gradually deplete, oil and gas drilling is increasingly shifting towards deeper and even ultra-deep formations. In these deep and ultra-deep areas, complex geological conditions increase the risk of dangerous incidents such as well blowouts. However, existing methods for training and utilizing models for drilling risk detection often fail to effectively and comprehensively capture and utilize the complex downhole environment, resulting in poor accuracy and a high likelihood of false alarms.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This manual provides a method and apparatus for determining drilling risks, which can accurately detect and identify drilling risks and reduce false alarm rates.

[0005] This manual provides a method for determining drilling risk, including:

[0006] Obtain the current target logging data of the target well, as well as the logging data group for the current time period;

[0007] Based on the target logging data, determine the current operating condition of the target well;

[0008] Based on the current working conditions of the target well, the logging data set for the current time period, and the preset geological-engineering pre-drilling assessment profile, the current target dynamic threshold of the target well is determined.

[0009] By processing target logging data using a pre-defined detection model, the corresponding target reconstruction error is obtained; the pre-defined detection model is a model trained by unsupervised learning based on a graph autoencoder.

[0010] Based on the target reconstruction error and the target dynamic threshold, determine whether there is a drilling risk in the target well.

[0011] In one embodiment, the target dynamic threshold of the target well is determined based on the current operating conditions of the target well, the logging data set for the current time period, and a preset geological-engineering pre-drilling assessment profile, including:

[0012] Based on the current working conditions of the target well, query the preset working condition threshold set and determine the matching target training set threshold.

[0013] By using the graph autoencoder in the preset detection model to process the logging data group of the current time period, the corresponding target test set threshold is determined;

[0014] Based on the target logging data, determine the current drill bit depth in the target well; and based on the current drill bit depth and the preset geological-engineering pre-drilling assessment profile, determine the corresponding target risk assessment value.

[0015] Based on the target training set threshold, the target test set threshold, and the target risk assessment value, the current target dynamic threshold of the target well is determined.

[0016] In one embodiment, the method further includes:

[0017] Acquire sample logging data; and construct a sample training set based on the sample logging data;

[0018] Based on the sample logging data, key characteristic parameters for drilling risks were determined through correlation analysis;

[0019] Based on drilling engineering rules and the rules of fluid mass conservation and pressure balance, the correlation between key characteristic parameters and drilling is determined; and based on the correlation, a correlation diagram of key characteristic parameters based on physical constraints for drilling risks is constructed.

[0020] Construct an initial detection model based on a graph autoencoder; wherein the graph autoencoder includes at least an encoder, and the encoder includes at least a sequentially connected GCN layer and a TCN layer;

[0021] The initial detection model is learned by using the correlation graph of key feature parameters so that the GCN layer learns the correlation constraints of different key feature parameters and obtains the intermediate detection model.

[0022] Using the sample training set, unsupervised learning is performed on the intermediate detection model to obtain a preset detection model that meets the requirements.

[0023] In one embodiment, the key characteristic parameters include at least one of the following: riser pressure, inlet flow rate, outlet flow rate, total pool volume, and casing pressure.

[0024] In one embodiment, the graph autoencoder further includes a decoder; wherein the decoder is connected to the encoder, and the decoder includes at least a TCN layer and a GCN layer connected in sequence.

[0025] In one embodiment, constructing a sample training set based on sample logging data includes:

[0026] From the sample logging data, the first type of sample logging data corresponding to the risk conditions and the second type of sample logging data corresponding to the normal conditions were selected.

[0027] A sample training set is constructed based on the second type of sample logging data; wherein, the second type of sample logging data is used to construct the test dataset.

[0028] In one embodiment, the method further includes:

[0029] Obtain target geological background data for the area where the target well is located, as well as historical drilling records of adjacent drilled wells.

[0030] Based on the target geological background data and the target historical drilling records, a set of risk accident factors associated with drilling risks is determined; wherein, the set of risk accident factors includes at least: geological factors and engineering factors;

[0031] Based on the aforementioned set of risk event factors, determine the fuzzy assessment rules for risk levels;

[0032] Based on the fuzzy risk level assessment rules, a comparison and judgment matrix for risk assessment is generated;

[0033] Based on the comparative judgment matrix of risk assessment, a corresponding pre-set geological-engineering pre-drilling assessment profile is constructed.

[0034] This manual also provides a method for determining drilling risk, including:

[0035] Obtain the current target logging data of the target well, as well as the logging data group for the current time period;

[0036] Based on the target logging data, determine the current operating condition of the target well;

[0037] Based on the current working conditions of the target well, the logging data set for the current time period, and the preset geological-engineering pre-drilling assessment profile, the current target dynamic threshold of the target well is determined.

[0038] By processing the target logging data using a pre-defined detection model, corresponding target result data is obtained; wherein, the pre-defined detection model includes at least an autoencoder;

[0039] Based on the target result data and target dynamic thresholds, determine whether the target well has drilling risks.

[0040] This specification also provides a device for determining drilling risk, comprising:

[0041] The acquisition module is used to acquire the current target logging data of the target well, as well as the logging data group for the current time period;

[0042] The first determining module is used to determine the current operating condition of the target well based on the target logging data;

[0043] The second determining module is used to determine the current target dynamic threshold of the target well based on the current working conditions of the target well, the logging data set of the current time period, and the preset geological-engineering pre-drilling assessment profile.

[0044] The processing module is used to obtain the corresponding target reconstruction error by processing the target logging data using a preset detection model; wherein, the preset detection model is a model trained by unsupervised learning based on a graph autoencoder;

[0045] The third determination module is used to determine whether there is a drilling risk in the target well based on the target reconstruction error and the target dynamic threshold.

[0046] This specification also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the relevant steps of the method for determining drilling risks.

[0047] Based on the drilling risk determination method and apparatus provided in this manual, before implementation, a graph autoencoder is introduced and used. Through unsupervised learning, a preset detection model based on physical constraints, adapted to drilling operation scenarios, and with good application results is trained. During implementation, the current target logging data and the logging data set for the current time period of the target well are first acquired. Based on the target logging data, the current working condition is determined. Then, based on the current working condition, the logging data set for the current time period, and the preset geological-engineering pre-drilling assessment profile, a target dynamic threshold is determined for the current target well, based on working condition constraints and considering the changes in data processed within the current time period before graph autoencoding, as well as the influence of macroscopic risk factors based on the depth domain. The preset detection model is used to process the target logging data to obtain the corresponding target reconstruction error. Then, based on the target reconstruction error and the target dynamic threshold, the existence of drilling risk is detected. This allows for the comprehensive capture and utilization of complex downhole environmental characteristics and, based on corresponding physical constraints, accurate detection and identification of drilling risks, effectively reducing the false alarm rate. Attached Figure Description

[0048] To more clearly illustrate the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating a method for determining drilling risk provided in one embodiment of this specification;

[0050] Figure 2 This is a schematic diagram illustrating one embodiment of the drilling risk determination method provided in the embodiments of this specification, applied in a scenario example.

[0051] Figure 3This is a schematic diagram illustrating one embodiment of the drilling risk determination method provided in the embodiments of this specification, applied in a scenario example.

[0052] Figure 4 This is a schematic diagram illustrating one embodiment of the drilling risk determination method provided in the embodiments of this specification, applied in a scenario example.

[0053] Figure 5 This is a schematic diagram illustrating one embodiment of the drilling risk determination method provided in the embodiments of this specification, applied in a scenario example.

[0054] Figure 6 This is a schematic diagram illustrating one embodiment of the drilling risk determination method provided in the embodiments of this specification, applied in a scenario example.

[0055] Figure 7 This is a schematic diagram illustrating one embodiment of the drilling risk determination method provided in the embodiments of this specification, applied in a scenario example.

[0056] Figure 8 This is a schematic diagram of the structural composition of a computer device provided in one embodiment of this specification;

[0057] Figure 9 This is a schematic diagram of the structural composition of a drilling risk determination device provided in one embodiment of this specification;

[0058] Figure 10 This is a schematic diagram illustrating one embodiment of the drilling risk determination method provided in the embodiments of this specification, applied in a scenario example.

[0059] Figure 11 This is a schematic diagram illustrating one embodiment of the drilling risk determination method provided in the embodiments of this specification, applied in a scenario example.

[0060] Figure 12 This is a schematic diagram illustrating one embodiment of the drilling risk determination method provided in the embodiments of this specification, applied in a scenario example. Detailed Implementation

[0061] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0062] It should be noted that the information and data related to users involved in the embodiments of this specification are all information and data authorized by the user or fully authorized by the relevant parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with relevant laws, regulations, and standards, and necessary confidentiality measures have been taken. They do not violate public order and good morals, and corresponding operation entry points are provided for users or relevant parties to choose to authorize or refuse.

[0063] It should also be noted that in the embodiments of this specification, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0064] See Figure 1 As shown in the embodiments of this specification, a method for determining drilling risk is provided. In specific implementation, this method may include the following:

[0065] S101: Obtain the current target logging data of the target well, as well as the logging data group for the current time period;

[0066] S102: Determine the current operating condition of the target well based on the target logging data;

[0067] S103: Based on the current working conditions of the target well, the logging data set for the current time period, and the preset geological-engineering pre-drilling assessment profile, determine the current target dynamic threshold of the target well.

[0068] S104: By processing the target logging data using a preset detection model, the corresponding target reconstruction error is obtained; wherein, the preset detection model is a model trained by unsupervised learning based on a graph autoencoder;

[0069] S105: Determine whether there is a drilling risk in the target well based on the target reconstruction error and the target dynamic threshold.

[0070] Specifically, the aforementioned target wells can be understood as wells currently undergoing drilling operations that require monitoring of drilling risks.

[0071] The aforementioned drilling risks may include one or more of the following: overflow risk, leakage risk, stuck pipe risk, etc.

[0072] The aforementioned target logging data can specifically refer to the logging data collected at the current time point corresponding to a specific time point.

[0073] The aforementioned logging data set for the current time period can specifically refer to the logging data stream collected and accumulated within the current time period, corresponding to that specific time period. For example, it could be all logging data related to the target well collected from the start of drilling operations up to the current time point.

[0074] In practice, the current operating condition of the target well can be determined by processing the target logging data using the operating condition identification model.

[0075] The current operating condition may specifically include at least one of the following: drilling, circulation, pump start-up, pump stop-up, etc. It should be noted that the operating conditions listed above are only illustrative. In actual implementation, depending on the specific circumstances, the current operating condition may also include other types of operating conditions besides those listed above. This instruction manual does not limit this.

[0076] The aforementioned working condition identification model can be a neural network model that is trained using sample data in advance through deep learning and can automatically identify and determine the corresponding working condition based on the input logging data.

[0077] The aforementioned pre-defined geological-engineering drilling assessment profile can be understood as a template profile that, from a macroscopic perspective, considers the combined impact of related risk factors, as well as the interrelationships between the well and adjacent wells, for specific drilling risks. Through fuzzy evaluation, it is constructed to estimate the corresponding drilling risk assessment score based on the depth domain (drill bit depth in the well). The specific construction method of the pre-defined geological-engineering drilling assessment profile will be explained separately later.

[0078] In practice, based on preset determination rules, the following steps can be taken: First, using the current operating conditions of the target well, the logging data set for the current time period, and the preset geological-engineering pre-drilling assessment profile, the training set threshold based on sample logging data and operating condition constraints, the test set threshold based on the actual data change characteristics of the logging data stream for the current time period, and the risk assessment value based on the impact of risk accident factors in the preset geological-engineering drilling assessment profile can be determined. Then, by combining the above-mentioned training set threshold, test set threshold, and risk assessment value, and other parameters based on different dimensions, the target dynamic threshold for the target well at the current time point can be accurately determined, which has high reference value and good real-time performance.

[0079] The aforementioned preset detection model can be understood as an algorithm model that uses an initial detection model containing at least a graph autoencoder, trained through unsupervised learning, to analyze and combine the temporal linkage changes of key feature parameters of multiple dimensions sensitive to drilling risks based on the logging data at a single input time point, and to judge and output the reconstruction error that can be used to indicate the drilling fluctuation situation by detecting whether there are abnormal changes.

[0080] Specifically, graph autoencoders (GAEs) can be understood as deep neural architectures that map nodes to a latent feature space and decode graph information from the latent representations. Specifically, GAEs can learn network embeddings or generate new graphs. The network embedding refers to the low-dimensional vector representation of nodes, which preserves the topological information of the nodes. Graph autoencoders learn network embeddings by extracting them using an encoder and using a decoder to ensure that the network embeddings preserve the graph's topological information (e.g., the PPMI matrix and adjacency matrix). For graph generation, given multiple graphs, GAEs can learn the generation distribution of graphs by encoding the graphs into hidden representations and then decoding the graph structure based on these hidden representations.

[0081] In practice, a preset detection model can be used to process the target logging data to obtain the target reconstruction error for the current target well. Then, based on the target reconstruction error and the target dynamic threshold, the target reconstruction error is compared to see if it exceeds the target dynamic threshold, yielding the corresponding comparison result. Based on the comparison result, if the target reconstruction error is determined to be greater than the target dynamic threshold, it can be determined that the target well has drilling risk. Conversely, based on the comparison result, if the target reconstruction error is determined to be less than or equal to the target dynamic threshold, it can be determined that the target well does not have drilling risk.

[0082] Based on the above embodiments, the target logging data at the current time point of the target well, as well as the logging data set for the current time period of the target well, can be acquired first. Simultaneously, considering the influence and constraints of multiple different dimensions, a target dynamic threshold for judging the drilling risk of the current target well can be determined. At the same time, using a pre-set detection model that has been trained through unsupervised learning and contains at least a graph autoencoder, the target logging data of the target well is processed to obtain the corresponding target reconstruction error. Furthermore, by combining the target reconstruction error and the target dynamic threshold, the existence of drilling risk in the target well can be accurately determined, effectively reducing detection errors and lowering the false alarm rate.

[0083] In some embodiments, see Figure 2 As shown, the above-mentioned autoencoder includes at least an encoder; wherein the encoder includes at least a GCN layer and a TCN layer connected in sequence.

[0084] Specifically, based on the GCN (Graph Convolutional Neural Network) layer, parameter constraints can be introduced through the graph structure, and the embedding of these constraints can be achieved using graph convolution methods. This demonstrates good performance in handling feature parameter associations and constraints, enabling the model to operate under the physical constraints of drilling operations, reducing detection errors and false alarm rates. Based on the TCN (Temporal Convolutional Neural Network) layer, corresponding temporal information can be extracted through causal dilation convolution stacking. This demonstrates good performance in handling long-term relationships in time series, enabling the model to accurately capture temporal dependencies in feature parameters.

[0085] Further reading Figure 2 As shown, the above-described graph autoencoder may further include a decoder; wherein the decoder is connected to the encoder. The decoder includes at least a TCN layer and a GCN layer connected in sequence. Furthermore, a low-dimensional layer may be connected between the encoder and the decoder.

[0086] When processing target logging data using a preset detection model, the target logging data can first be acquired and converted into a corresponding time-series-based raw sequence based on temporal characteristics. The raw sequence is then input into the encoder in the graph autoencoder for encoding and dimensionality reduction, and then into the decoder for decoding. Finally, it is input into the graph autoencoder for reconstruction to obtain the corresponding reconstructed sequence. Based on the reconstructed sequence and the raw sequence, the corresponding reconstruction error is calculated and output.

[0087] In some embodiments, after acquiring the current target logging data of the target well, the method may further include: preprocessing the target logging data to obtain preprocessed target logging data. The preprocessing may include one or more of the following: outlier detection, missing value handling, interpolation, data smoothing, and normalization. Preprocessing makes the target logging data more suitable for subsequent processing.

[0088] In some embodiments, see Figure 3 As shown, the target dynamic threshold for the target well is determined based on the current operating conditions of the target well, the logging data set for the current time period, and the preset geological-engineering pre-drilling assessment profile. In specific implementation, this may include the following:

[0089] S1: Based on the current working condition of the target well, query the preset working condition threshold set and determine the matching target training set threshold.

[0090] S2: By using the graph autoencoder in the preset detection model to process the logging data group of the current time period, the corresponding target test set threshold is determined;

[0091] S3: Based on the target logging data, determine the current drill bit depth in the target well; and based on the current drill bit depth and the preset geological-engineering pre-drilling assessment profile, determine the corresponding target risk assessment value;

[0092] S4: Determine the current target dynamic threshold of the target well based on the target training set threshold, the target test set threshold, and the target risk assessment value.

[0093] The preset operating condition threshold set is a set of thresholds determined in advance based on operating condition constraints through learning and statistical analysis of a large amount of sample logging data. Specifically, the preset operating condition threshold set may include multiple training set thresholds, each training set threshold corresponding to at least one operating condition.

[0094] In practice, the logging data set for the current time period can be converted into a corresponding data stream. Then, using a sliding time window, logging data from multiple consecutive time points is extracted from the data stream. This data is then processed using a graph autoencoder in a pre-defined detection model to obtain the processing results. Based on these results, statistics are performed to determine a target test set threshold that characterizes the current data fluctuation trend and has strong real-time performance for the current time point. This approach yields a test set threshold that simultaneously matches the knowledge learned and mastered by the graph autoencoder in the pre-defined detection model and the local short-term fluctuation trend of the logging data within the current time period.

[0095] For example, a sliding window with a window length of L=60 can be used to operate on the logging data group of the current time period, and then the data after operation can be processed using a preset detection model. The standard coefficient n=1 is set, and statistics are performed to calculate the real-time test set threshold.

[0096] In practice, the weight coefficients corresponding to the target training set threshold and the target test set threshold can be determined according to the preset determination rules. Then, the target training set threshold and the target risk assessment value are combined to determine the corresponding reference assessment items. Based on the corresponding weight coefficients, the target training set threshold, the target test set threshold, and the reference assessment items are used together for weighted calculation to determine the final target dynamic threshold.

[0097] Specifically, for example, the current target dynamic threshold of the target well can be determined according to the following formula based on the target training set threshold, the target test set threshold, and the target risk assessment value:

[0098] T n =W y L a +W z G a +L a(1-R(H))

[0099] Among them, T n Let n represent the current time point, and L be the target dynamic threshold. a G is the threshold of the target training set. a R(H) is the target test set threshold, R(H) is the target risk assessment value (e.g., target risk assessment score), H is the current drill bit depth in the target well, and L is the target test set threshold. a (1-R(H)) is the reference evaluation item, W y W z These are the first weight coefficient and the second weight coefficient, corresponding to the target training set threshold and the target test set threshold, respectively. Specifically, the first and second weight coefficients can be determined in advance using a large amount of sample logging data, through data statistics and relevant big data analysis.

[0100] Based on the above embodiments, the target dynamic threshold, which is suitable for the drilling risk assessment of the current target well and has strong real-time performance, can be accurately determined by comprehensively and fully considering the influence and constraints of multiple different dimensions, based on the current working conditions of the target well, the logging data group of the current time period, and the preset geological-engineering pre-drilling assessment profile.

[0101] In some embodiments, see Figure 4 As shown, the above method may also include the following in its specific implementation:

[0102] S1: Obtain sample logging data; and construct a sample training set based on the sample logging data;

[0103] S2: Based on the sample logging data, key characteristic parameters for drilling risks are determined through correlation analysis;

[0104] S3: Based on drilling engineering rules and the rules of fluid mass conservation and pressure balance, determine the correlation between key characteristic parameters and drilling; and based on the correlation, construct a correlation diagram of key characteristic parameters based on physical constraints for drilling risks;

[0105] S4: Construct an initial detection model based on a graph autoencoder; wherein the graph autoencoder includes at least an encoder, and the encoder includes at least a sequentially connected GCN layer and a TCN layer;

[0106] S5: Use the key feature parameter correlation graph to perform prior learning on the initial detection model so that the GCN layer learns the correlation constraints of different key feature parameters and obtains the intermediate detection model.

[0107] S6: Using the sample training set, perform unsupervised learning on the intermediate detection model to obtain a preset detection model that meets the requirements.

[0108] The aforementioned training set contains multiple sample logging data, and the sample logging data is unlabeled, so that the model can be used for unsupervised learning using the aforementioned training set.

[0109] Based on the above embodiments, the data characteristics of sample data in drilling construction scenarios and the engineering mechanisms in the drilling process can be effectively considered and utilized. Through unsupervised learning, the changing patterns of logging data under normal working conditions can be learned, and a preset detection model adapted to drilling construction scenarios can be constructed.

[0110] In some embodiments, specific implementations may refer to Figure 5 As shown, a correlation analysis was conducted on drilling risk based on the sample logging data, and the corresponding correlation analysis results were obtained. Based on the correlation analysis results and combined with the engineering constraints in drilling operations, several characteristic parameters with high sensitivity to drilling risk were selected as key characteristic parameters.

[0111] In some embodiments, the key characteristic parameters may specifically include at least one of the following: riser pressure (or pump pressure), inlet flow rate, outlet flow rate, total tank volume, casing pressure, etc.

[0112] Of course, it should be noted that the key characteristic parameters listed above are only illustrative. In actual implementation, depending on the specific circumstances and processing requirements, the key characteristic parameters may also include other phase-adaptive parameters such as total hydrocarbon content, in addition to those listed above.

[0113] Specifically, riser pressure is taken into consideration because pressure changes are a crucial indicator for monitoring downhole dynamics during drilling operations. When a blow occurs, the drilling fluid begins to flow freely, disrupting the equilibrium of the hydrostatic column. Due to the upward movement of the drilling fluid, the hydrostatic column height decreases, and the bottomhole pressure drops accordingly. However, inside the drill string, the impact of fluid intrusion is relatively small, and the hydrostatic pressure and circulation pressure loss remain almost constant. Therefore, the decrease in bottomhole pressure leads to a decrease in riser pressure, a change that is significant for predicting blows.

[0114] Changes in inlet and outlet flow rates can also reflect the occurrence of overflow risk. Under normal circumstances, inlet and outlet flow rates should remain relatively stable. However, when an overflow occurs, the outlet flow rate will increase without an increase in the inlet flow rate, which is an important characteristic for identifying an overflow.

[0115] Total pool volume, as a key parameter describing drilling fluid reserves, also significantly impacts blowout prediction. As drilling fluid is returned, the total pool volume gradually increases. In the case of oil-water intrusion, the total pool volume increases accordingly with the increase in outlet flow rate; similarly, in the case of gas intrusion, the total pool volume increases due to the expansion of gas bubbles and the resulting increase in outlet flow rate. These changes become apparent shortly after oil-water intrusion and gas intrusion, providing crucial clues for blowout prediction.

[0116] After identifying the key feature parameters, the sample logging data can be cleaned in a targeted manner based on the selected key feature parameters to obtain sample logging data with smaller errors and better results. This sample logging data can then be used to effectively eliminate relevant data errors and improve the model training accuracy.

[0117] In some embodiments, specific implementations may refer to Figure 6 As shown, the correlation (e.g., causal relationship) of key characteristic parameters with respect to drilling can be determined based on drilling engineering rules (e.g., drilling engineering principles) and fluid mass conservation and pressure balance rules; and based on the correlation, a correlation diagram of key characteristic parameters with respect to drilling risk based on physical constraints can be constructed.

[0118] Specifically, for example, we can first determine the relationship between the casing pressure (which can be abbreviated as casing pressure) and the pump pressure (or, riser pressure) as a pressure drop formula; the relationship between the pump pressure and the inlet flow rate as a pump start-stop constraint; the relationship between the inlet flow rate and the outlet flow rate as a flow conservation formula; the relationship between the outlet flow rate and the total tank volume as a flow conservation formula; and the relationship between the pump pressure and the total tank volume as a pump start-stop constraint. Then, we set up nodes corresponding to the above key feature parameters and connect two nodes with a relationship using edges. Based on the specific relationships between the key feature parameters, we label the edges between the corresponding two nodes, thereby embedding the physical mechanism and specific operating condition constraints to obtain a key feature parameter relationship graph based on physical constraints.

[0119] In practical implementation, the key feature parameter correlation graph can be used as prior data and input into the GCN layer of the graph autoencoder in the initial detection model for prior learning. This achieves graph embedding of key feature parameters, thereby strengthening the GCN layer's constraints and linkages between key feature parameters by introducing explicit physical constraints. This allows the detection model to better distinguish and filter heterogeneous errors caused by switching between normal operating conditions and normal fluctuations in data within normal operating conditions, enhancing model accuracy and reducing false alarm rates.

[0120] In specific implementation, after completing the prior learning of the GCN layer for the graph autoencoder and obtaining the intermediate detection model, a specified number (e.g., 500) of consecutive time-point sample logging data can be randomly and continuously extracted from the sample training set. These samples are then connected according to the drilling process sequence to obtain the corresponding sample logging data sequence. This sample logging data sequence is then input into the intermediate detection model. Furthermore, the GCN layer of the encoder in the graph autoencoder can first form a chain propagation path for the sample logging data sequence to learn and capture the gradual change patterns of key feature parameters with drilling, obtaining the corresponding change information. Then, the graph... In the autoencoder, the TCN layer of the encoder exponentially expands the receptive field using dilated causal convolution, enabling the learning and perception of multi-scale time series of key feature parameters, thus obtaining the corresponding temporal information. This change information and temporal information are then combined with the sample logging data sequence and subjected to dimensionality reduction by a low-dimensional layer before being fed into the decoder for decoding, resulting in the reconstructed sequence. The corresponding reconstruction error is then calculated based on the reconstructed sequence, and the loss function value is calculated based on the reconstruction error. Finally, based on the loss function value, the model parameters are iteratively optimized for the current round to obtain the updated detection model, completing one round of unsupervised learning.

[0121] Following the above method, after completing multiple rounds of unsupervised learning, when the calculated loss function value is less than or equal to the preset loss threshold, the model training ends, and the current detection model is determined as the preset detection model that meets the requirements.

[0122] In some embodiments, the above-described construction of a sample training set based on sample logging data may include the following:

[0123] S1: Select the first type of sample logging data corresponding to the risk conditions and the second type of sample logging data corresponding to the normal conditions from the sample logging data;

[0124] S2: Construct a sample training set based on the second type of sample logging data; wherein, the second type of sample logging data is used to construct a test dataset.

[0125] Specifically, considering that the number of first-class sample logging data corresponding to risk conditions is relatively sparse compared to the number of second-class sample logging data corresponding to normal conditions in the normally acquired sample logging data, it is possible to construct a sample training set using only the second-class sample logging data. This training set can then be used for unsupervised learning of the model, allowing it to focus on the feature representation of logging data under normal conditions and the fluctuation patterns of related features under normal conditions. In this way, the trained model can automatically identify drilling risks under risk conditions by detecting and distinguishing fluctuations in related features that differ from those under normal conditions.

[0126] In specific implementation, the method may further include: constructing a test dataset based on the first type of sample logging data.

[0127] Accordingly, the above-mentioned unsupervised learning of the intermediate detection model using the sample training set to obtain a preset detection model that meets the requirements can be implemented as follows: using the sample training set to perform unsupervised learning of the intermediate detection model, so that the model can deeply learn the characteristic fluctuation and change law of logging data under normal working conditions, and obtain a candidate model; using the test dataset to test the candidate model and obtain the corresponding test results; judging whether the candidate model meets the requirements based on the test results; and determining the current candidate model as the preset detection model if the candidate model meets the requirements.

[0128] Conversely, if the candidate model is determined to be unsuitable, the model parameters in the candidate model can be adjusted based on the test results to obtain an adjusted candidate model. The adjusted candidate model can then be used as an intermediate detection model. Unsupervised learning can be continued on the intermediate detection model using the sample training set until a preset detection model that meets the requirements is obtained.

[0129] The above-mentioned construction of a test dataset based on the first type of sample logging data can specifically include: querying sample logging data to determine sample logging data of adjacent wells in normal operating conditions corresponding to the well segments of the first type of sample logging data, as auxiliary sample logging data; and combining the first type of sample logging data and the auxiliary sample logging data to obtain the test dataset. The test dataset can include multiple test datasets, each consisting of the first type of sample logging data and the corresponding auxiliary sample logging data, with the ratio of the first type of sample logging data to the auxiliary sample logging data in a test dataset satisfying a preset ratio parameter. For example, the ratio of the first type of sample logging data to the auxiliary sample logging data in a test dataset is greater than or equal to 4:6. This allows the model to be tested and adjusted using the test dataset, enabling it to accurately distinguish the characteristic fluctuations in logging data with drilling risks under risky operating conditions; and also allows the model to predict potential drilling risks in advance based on the continuity characteristics between logging data under normal operating conditions and logging data under risky operating conditions.

[0130] In some embodiments, the above-described construction of a sample training set based on sample logging data may include the following:

[0131] S1: Select the first type of sample logging data corresponding to the risk conditions and the second type of sample logging data corresponding to the normal conditions from the sample logging data;

[0132] S2: Randomly split the second type of sample logging data into multiple second type of sample logging data groups;

[0133] S3: Combine the first type of sample logging data with multiple second type sample logging data groups to obtain multiple sample training data groups; and combine the multiple sample training data groups to obtain a sample training set; wherein the proportion of the first type sample logging data included in each sample training data group is greater than a preset proportion threshold. The preset proportion threshold can be 60%.

[0134] Specifically, considering that the first type of sample logging data corresponding to risky working conditions is relatively sparse in the normally acquired sample logging data, the more numerous second type sample logging data is split into multiple second type sample logging data groups, and then combined with the first type sample logging data. This ensures that each sample training data group contains at least a corresponding number of first type sample logging data. In this way, when using the above sample training data groups to perform unsupervised learning on the model, it can avoid the model becoming lazy, which would prevent it from accurately identifying and discovering drilling risks.

[0135] In some embodiments, constructing a sample training set based on sample logging data may further include: selecting sample logging data without drilling risk from the sample logging data to construct a normal dataset; simultaneously, selecting sample logging data with drilling risk as abnormal data from the sample logging data; querying the corresponding sample drilling logs to determine the risk well section corresponding to the abnormal data, and determining the normal well section adjacent to the risk well section; obtaining a first specified proportion (e.g., 40%) of sample logging data for the normal well section, and combining it with a second specified proportion (e.g., 60%) of sample logging data for the risk well section in chronological order to obtain the corresponding risk dataset; combining the normal dataset and the risk dataset; and combining the multiple sample training data groups to obtain the sample training set.

[0136] Based on the above embodiments, a sample training set suitable for drilling risk detection and unsupervised learning can be constructed by making full use of the larger number of second-type sample logging data and the smaller number of first-type sample logging data.

[0137] After constructing the sample training set, it can be further divided into a first training set and a first test set according to the cross-well partitioning rules; wherein the wells corresponding to the first training set are different from those corresponding to the first test set.

[0138] In this way, the model can be trained using the first training set and then improved using the first test set, so that the model has better generalization ability.

[0139] In some embodiments, see Figure 7As shown, in specific implementations, the method may also include the following:

[0140] S1: Obtain target geological background data for the area where the target well is located, as well as the target historical drilling records of adjacent drilled wells;

[0141] S2: Based on the target geological background data and the target historical drilling records, determine the set of risk accident factors associated with drilling risks; wherein, the set of risk accident factors includes at least: geological factors and engineering factors;

[0142] S3: Determine the fuzzy assessment rules for risk levels based on the aforementioned set of risk event factors;

[0143] S4: Generate a comparison judgment matrix for risk assessment based on the fuzzy risk level assessment rules;

[0144] S5: Based on the comparative judgment matrix of risk assessment, construct the corresponding preset geological-engineering pre-drilling assessment profile.

[0145] The geological factors include at least one of the following: formation lithology, formation pressure coefficient, geological stability (or pressure uncertainty), etc. The engineering factors include at least one of the following: the difference between predicted ECD (i.e., drilling fluid circulating equivalent density) and formation pressure, the severity of risks occurring in adjacent wells (or adjacent well risk coefficient), etc.

[0146] Specifically, in this embodiment, the geological factors used may include: formation pressure coefficient and pressure uncertainty; the engineering factors used may include: adjacent well risk coefficient, predicted ECD and formation pressure difference, etc. In practice, expert experience can be combined with big data analysis using target geological background data and target historical drilling records to identify multiple image factors that have a significant impact on drilling risk in the target well area, which can be used as risk accident factors; by combining the above risk accident factors, a set of risk accident factors associated with drilling risk can be constructed.

[0147] When determining specific fuzzy assessment rules for risk levels, drilling accident records from adjacent drilled wells can be combined with historical drilling accident records to perform machine learning on the correlation between specific risk accident factors (including geological and engineering factors) and drilling accidents in the risk accident factor set. Based on the machine learning results, sub-assessment rules for different risk levels based on different risk accident factors, as well as coupling weight matrices between different risk levels, can be determined.

[0148] Each sub-assessment rule corresponds to a risk event factor and includes a risk level classification method based on that risk event factor. Specifically, each sub-assessment rule contains at least three risk levels: "low risk," "medium risk," and "high risk," along with three data value ranges corresponding to each of these risk levels for that risk event factor. Combining these multiple self-assessment rules yields the corresponding fuzzy risk level assessment rules. See Table 1 for details on the fuzzy risk level assessment.

[0149] Table 1

[0150]

[0151] The aforementioned coupling weight matrix can be, for example, W = [1 2 5]. T .

[0152] After obtaining the fuzzy risk assessment rules, based on these rules and using the tomography method, drilling accident records from adjacent drilled wells, along with other historical drilling accident records, and expert experience, are used to score each risk factor. Then, based on the scores of each risk factor, pairwise comparisons are made to determine the relative contribution ratio of any two risk factors in indicating drilling risk. Based on these relative contribution ratios, a corresponding risk assessment comparison matrix is ​​constructed. For example, Among them, a ij This represents the relative contribution ratio of risk event factor numbered i to risk event factor numbered j.

[0153] After obtaining the comparison judgment matrix for risk assessment, the formula Mw = l can be solved based on this matrix. max w, determine the corresponding auxiliary matrix V M =w; where, l max This is the largest eigenvalue of the comparison judgment matrix. For example, the determined auxiliary matrix could be V. M =[0.653,0.252,0.104,0.705] T .

[0154] By combining auxiliary matrices and coupling weight matrices, a pre-defined geological-engineering pre-drilling assessment profile is constructed. Specifically, the pre-defined geological-engineering pre-drilling assessment profile can be constructed according to the following formula:

[0155] B h =V M ×S h ×W

[0156] Among them, B hThe risk assessment value is determined based on a pre-set geological-engineering pre-drilling assessment profile, where h is the drill bit depth and V is the vertical depth. M S is the auxiliary matrix, W is the coupling weight matrix, and S is the coupling weight matrix. h This is a fuzzy risk level matrix determined based on the fuzzy risk level assessment rules and the data values ​​of multiple risk event factors at point h.

[0157] In practice, after obtaining the comparison judgment matrix for risk assessment, a consistency check can be performed on the matrix. If the consistency check passes, the current comparison judgment matrix is ​​considered to have high representativeness and reference value, meeting the usage requirements. Conversely, if the consistency check fails, the risk level fuzzy assessment rules and / or corresponding elements in the comparison judgment matrix can be adjusted in a targeted manner until the consistency check passes.

[0158] When performing consistency checks, you can first calculate the largest eigenvalue l of the comparison judgment matrix. max Then, the corresponding consistency parameter C is determined based on the largest eigenvalue. I Simultaneously, based on the number of risk event factors (e.g., n) in the risk event factor set, the corresponding consistency index R is determined by querying the analytic hierarchy process table. I Next, calculate the ratio between the consistency parameter and the consistency index to obtain the corresponding consistency ratio C. R The system checks whether the consistency ratio is less than or equal to a preset ratio threshold (e.g., 0.1). If the consistency ratio is determined to be less than or equal to the preset ratio threshold, the consistency check is considered to have passed; otherwise, the consistency check is considered to have failed.

[0159] Specifically, the consistency ratio can be calculated using the following formula:

[0160] C I =(l max -n) / (n-1)

[0161]

[0162] Among them, C R C represents the consistency ratio. I For consistency parameters, l max To compare the largest eigenvalue of the judgment matrix, where n is the number of risk factors in the risk factor set, R I This serves as a consistency indicator.

[0163] Based on the above embodiments, a pre-set geological-engineering pre-drilling assessment profile with good performance based on the depth domain can be constructed, which is adapted to drilling construction scenarios, fully considers the macroscopic effects of various risk and accident factors, and is aligned with the temporal processing of the model.

[0164] In some embodiments, the above-mentioned determination of the corresponding target risk assessment value based on the current drill bit depth and the preset geological-engineering pre-drilling assessment profile may include the following:

[0165] S1: Determine the data values ​​of the current risk factors based on the current drill bit depth;

[0166] S2: Based on the fuzzy risk level assessment rules and the data values ​​of current risk event factors, determine the risk level based on each risk event factor; and based on the risk level, determine the current fuzzy risk level matrix;

[0167] S3: Based on the preset geological-engineering pre-drilling assessment profile, the corresponding target risk assessment value is calculated using the current fuzzy risk level matrix.

[0168] In practice, the data values ​​of each risk factor can be determined based on the current drill bit depth and the historical drilling records of adjacent drilled wells.

[0169] Specifically, for example, when determining the risk coefficient of a current adjacent well, one can first identify historical risk locations in the target area where the target well is located, based on the target's historical drilling records; then, obtain the distance parameters of these historical risk locations relative to the current drill bit depth of the target well; next, determine the risk score for each historical risk location based on the target's historical drilling records; and finally, determine the distance weight of each historical risk location based on its distance parameters. See Table 2 for reference.

[0170] Table 2

[0171]

[0172] Then, based on the risk scores of each historical risk location and the corresponding distance weights, the data value of the current adjacent well risk coefficient corresponding to the current drill bit depth position of the target well is calculated.

[0173] Specifically, the current risk coefficient of adjacent wells can be calculated as follows:

[0174]

[0175] Among them, w iLet l be the distance weight of the historical risk location point numbered i. i Let i be the distance parameter of the historical risk location point, n be the total number of historical risk locations, and r be the distance parameter of the location point. i For the risk score of historical risk location point numbered i, r a This represents the current risk coefficient value for adjacent wells.

[0176] Similarly, based on the current drill bit depth, the current formation pressure coefficient, the current pressure uncertainty, and the difference between the current predicted ECD and formation pressure can also be determined. For example, when the current drill bit depth is 2500 meters and 3930 meters, the determined data values ​​for each current risk factor can be found in Table 3.

[0177] Table 3

[0178]

[0179] Furthermore, based on the current fuzzy risk assessment rules and the current data values ​​of risk event factors, the risk level for each risk event factor can be determined; and based on the risk level, the current fuzzy risk level matrix can be determined.

[0180] For example, when the drill bit reaches a vertical depth of 2500 meters, the determined fuzzy risk level matrix can be represented as follows: In this context, the first row vector [1 0 0] indicates that the risk level determined based on the risk coefficient of adjacent wells according to the fuzzy risk level assessment rule is low risk; the second row vector [1 0 0] indicates that the risk level determined based on the formation pressure coefficient according to the fuzzy risk level assessment rule is low risk; the third row vector [0 0 1] indicates that the risk level determined based on the pressure uncertainty according to the fuzzy risk level assessment rule is high risk; and the fourth row vector [1 0 0] indicates that the risk level determined based on the difference between the predicted ECD and the formation pressure according to the fuzzy risk level assessment rule is low risk.

[0181] Furthermore, using the aforementioned fuzzy risk level matrix, the target risk assessment value at a drill bit depth of 2500 meters can be calculated according to the following formula: B 2500 =V M ×S 2500 ×W = 2.13.

[0182] Based on the above embodiments, the knowledge and information contained in the preset geological-engineering pre-drilling assessment profile can be fully utilized to determine a risk assessment value with high reference value from a macro perspective and based on the depth domain.

[0183] As can be seen from the above, the drilling risk determination method provided in the embodiments of this specification introduces and uses a graph autoencoder before implementation. Through unsupervised learning, a preset detection model based on physical constraints and with good application performance is trained. Furthermore, a preset set of working condition thresholds based on working condition constraints and a preset geological-engineering pre-drilling assessment profile that fully considers the influence of various risk factors are determined and constructed. In specific implementation, the current target logging data and the logging data set for the current time period of the target well can be obtained first. Based on the target logging data, the current working condition is determined. Then, based on the current working condition, the logging data set for the current time period, and the preset geological-engineering pre-drilling assessment profile, a target dynamic threshold is determined for the current target well, based on working condition constraints and simultaneously considering the changes in the data processed within the current time period before graph autoencoding, as well as the influence of macroscopic risk factors based on the depth domain. The preset detection model is used to process the target logging data to obtain the corresponding target reconstruction error. Finally, based on the target reconstruction error and the target dynamic threshold, the existence of drilling risk is detected. This allows for the comprehensive capture and utilization of complex downhole environmental characteristics, and based on corresponding physical constraints, accurate detection and identification of drilling risks, effectively reducing false alarm rates.

[0184] This manual also provides another method for determining drilling risk, which, in practice, may include the following:

[0185] S1: Obtain the current target logging data of the target well, as well as the logging data group for the current time period;

[0186] S2: Determine the current operating condition of the target well based on the target logging data;

[0187] S3: Based on the current working conditions of the target well, the logging data set for the current time period, and the preset geological-engineering pre-drilling assessment profile, determine the current target dynamic threshold of the target well;

[0188] S4: Using a preset detection model to process the target logging data, obtain the corresponding target result data; wherein, the preset detection model includes at least an autoencoder;

[0189] S5: Based on the target result data and target dynamic threshold, determine whether there is a drilling risk in the target well.

[0190] The aforementioned autoencoder (AE) can be understood as a type of artificial neural network (ANN) structure used in semi-supervised and unsupervised learning. Its function is to perform representation learning on the input information by using the input information as the learning target.

[0191] Specifically, the aforementioned autoencoder can be a graph autoencoder or other suitable autoencoders besides graph autoencoders.

[0192] Accordingly, the aforementioned target result data can be matched with the autoencoder used. Specifically, the aforementioned target result data can be reconstruction error, probability values, etc.

[0193] Based on the above embodiments, drilling risks can be accurately detected and identified, and false alarms can be reduced, by using dynamic thresholds in combination with a preset detection model that includes at least an autoencoder.

[0194] This specification provides an embodiment of a computer device, see below. Figure 8 As shown. The computer device includes a network communication port 801, a processor 802, and a memory 803. These structures are connected by internal cables so that they can perform specific data interaction.

[0195] Specifically, the network communication port 801 can be used to acquire the current target logging data of the target well, as well as the logging data group for the current time period.

[0196] The processor 802 can be specifically used to determine the current operating condition of the target well based on the target logging data; determine the current target dynamic threshold of the target well based on the current operating condition of the target well, the logging data set of the current time period, and the preset geological-engineering pre-drilling assessment profile; obtain the corresponding target reconstruction error by processing the target logging data using a preset detection model; wherein the preset detection model is a model trained by unsupervised learning based on a graph autoencoder; and determine whether there is drilling risk in the target well based on the target reconstruction error and the target dynamic threshold.

[0197] The memory 803 can be used to store corresponding instruction programs, as well as related data such as preset geological-engineering pre-drilling assessment profiles and preset detection models.

[0198] Based on the above method, the relevant structural performance of computer equipment can be effectively utilized to improve the data processing speed of electronic equipment and efficiently realize the data processing for determining drilling risks.

[0199] In this embodiment, the network communication port 801 can be a virtual port bound to different communication protocols, thereby enabling the sending or receiving of different data. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. Furthermore, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM or CDMA; it can also be a Wi-Fi chip; or it can be a Bluetooth chip.

[0200] In this embodiment, the processor 802 can be implemented in any suitable manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. This specification is not limiting.

[0201] In this embodiment, the memory 803 may include multiple layers. In a digital system, anything that can store binary data can be a memory. In an integrated circuit, a circuit with storage function but no physical form is also called a memory, such as RAM, FIFO, etc. In a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0202] This specification also provides a computer-readable storage medium based on the above-described method for determining drilling risk. The computer-readable storage medium stores computer program instructions that, when executed, perform the following: acquiring current target logging data and logging data sets for the current time period of the target well; determining the current operating condition of the target well based on the target logging data; determining the current target dynamic threshold of the target well based on the current operating condition of the target well, the logging data sets for the current time period, and a preset geological-engineering pre-drilling assessment profile; obtaining the corresponding target reconstruction error by processing the target logging data using a preset detection model; wherein the preset detection model is a model trained through unsupervised learning based on a graph autoencoder; and determining whether the target well has drilling risk based on the target reconstruction error and the target dynamic threshold.

[0203] In this embodiment, the storage medium includes, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), cache, hard disk drive (HDD), or memory card. The memory can be used to store computer program instructions. The network communication unit can be an interface configured according to standards specified in the communication protocol for network connection communication.

[0204] In this embodiment, the specific functions and effects implemented by the program instructions stored in the computer-readable storage medium can be explained in comparison with other embodiments, and will not be repeated here.

[0205] This specification also provides a computer program product, comprising at least a computer program, which, when executed by a processor, implements the following method steps: acquiring the current target logging data of the target well and the logging data set for the current time period; determining the current operating condition of the target well based on the target logging data; determining the current target dynamic threshold of the target well based on the current operating condition of the target well, the logging data set for the current time period, and a preset geological-engineering pre-drilling assessment profile; obtaining the corresponding target reconstruction error by processing the target logging data using a preset detection model; wherein the preset detection model is a model trained by unsupervised learning based on a graph autoencoder; and determining whether there is a drilling risk in the target well based on the target reconstruction error and the target dynamic threshold.

[0206] See Figure 9 As shown in the embodiments of this specification, a drilling risk determination device is also provided, which may specifically include the following structural modules:

[0207] The acquisition module 901 can be used to acquire the current target logging data of the target well, as well as the logging data group for the current time period;

[0208] The first determining module 902 can be used to determine the current working condition of the target well based on the target logging data;

[0209] The second determining module 903 can be used to determine the current target dynamic threshold of the target well based on the current working conditions of the target well, the logging data group of the current time period, and the preset geological-engineering pre-drilling assessment profile.

[0210] The processing module 904 can be used to obtain the corresponding target reconstruction error by processing the target logging data using a preset detection model; wherein, the preset detection model is a model trained by unsupervised learning based on a graph autoencoder.

[0211] The third determination module 905 can be used to determine whether there is a drilling risk in the target well based on the target reconstruction error and the target dynamic threshold.

[0212] In some embodiments, when the second determining module 903 is specifically implemented, it can determine the current target dynamic threshold of the target well according to the current working conditions of the target well, the logging data set of the current time period, and the preset geological-engineering pre-drilling assessment profile in the following manner: Based on the current working conditions of the target well, query the preset working condition threshold set to determine the matching target training set threshold; process the logging data set of the current time period using the graph autoencoder in the preset detection model to determine the corresponding target test set threshold; determine the current drill bit depth in the target well based on the target logging data; and determine the corresponding target risk assessment value based on the current drill bit depth and the preset geological-engineering pre-drilling assessment profile; and determine the current target dynamic threshold of the target well based on the target training set threshold, the target test set threshold, and the target risk assessment value.

[0213] In some embodiments, the device may also be used to: acquire sample logging data; construct a sample training set based on the sample logging data; determine key feature parameters for drilling risks through correlation analysis based on the sample logging data; determine the correlation relationship between the key feature parameters and drilling based on drilling engineering rules and fluid mass conservation and pressure balance rules; construct a key feature parameter correlation graph for drilling risks based on physical constraints based on the correlation relationship; construct an initial detection model based on a graph autoencoder; wherein the graph autoencoder includes at least an encoder, and the encoder includes at least a sequentially connected GCN layer and a TCN layer; perform prior learning on the initial detection model using the key feature parameter correlation graph, so that the GCN layer learns the correlation constraints of different key feature parameters to obtain an intermediate detection model; and perform unsupervised learning on the intermediate detection model using the sample training set to obtain a preset detection model that meets the requirements.

[0214] In some embodiments, the key characteristic parameters may specifically include at least one of the following: riser pressure, inlet flow rate, outlet flow rate, total pool volume, casing pressure, etc.

[0215] In some embodiments, the graph autoencoder further includes a decoder; wherein the decoder is connected to the encoder, and the decoder includes at least a TCN layer and a GCN layer connected in sequence.

[0216] In some embodiments, when the device is specifically implemented, a sample training set can be constructed based on the sample logging data in the following manner: a first type of sample logging data corresponding to risk conditions and a second type of sample logging data corresponding to normal conditions are selected from the sample logging data; a sample training set is constructed based on the second type of sample logging data; wherein, the second type of sample logging data is used to construct a test dataset.

[0217] In some embodiments, the device may also be used to: acquire target geological background data of the area where the target well is located, and target historical drilling records of adjacent drilled wells; determine a set of risk accident factors associated with drilling risk based on the target geological background data and target historical drilling records; wherein the set of risk accident factors includes at least geological factors and engineering factors; determine fuzzy risk level assessment rules based on the set of risk accident factors; generate a risk assessment comparison judgment matrix based on the risk level fuzzy assessment rules; and construct a corresponding preset geological-engineering pre-drilling assessment profile based on the risk assessment comparison judgment matrix.

[0218] It should be noted that the units, devices, or modules described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above devices are described by dividing them into various modules according to their functions. Of course, in implementing this specification, the functions of each module can be implemented in one or more software and / or hardware, or the module that implements the same function can be implemented by a combination of multiple sub-modules or sub-units, etc. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection between the devices or units shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0219] As can be seen from the above, the drilling risk determination device provided in the embodiments of this specification can comprehensively capture and utilize the complex environmental characteristics downhole, and accurately detect and identify drilling risks based on corresponding physical constraints, effectively reducing the false alarm rate.

[0220] In a specific scenario example, the drilling risk determination method provided in this manual can be applied to achieve drilling risk detection by combining geological and engineering information. The specific implementation process can be found below.

[0221] In this scenario example, considering that when using existing methods to monitor drilling risks using model algorithms, the common artificial intelligence algorithms used often require a large number of risk incident samples for model training. However, in the field of drilling engineering, the samples of drilling risks such as overflows (e.g., the first type of sample logging data) are inevitably sparse compared to the samples of normal operating conditions (e.g., the second type of sample logging data), resulting in severe sample imbalance, which will prevent the model from accurately identifying overflow risks.

[0222] To address the aforementioned problems and their root causes, this paper considers introducing both unsupervised model training and dynamic thresholding methods in drilling risk monitoring. Unsupervised model training refers to training a model that learns the characteristics of the samples themselves without requiring input sample labels. Dynamic thresholding methods involve dynamically adjusting thresholds in the decision-making layer of the unsupervised intelligent model using statistical or physical constraints to optimize alarm strategies. Furthermore, an autoencoder is introduced into the model during unsupervised training, and the autoencoder is constrained and optimized using geological-engineering prior embeddings. This improves the efficiency of model deployment while ensuring the accuracy of risk diagnosis, enabling rapid training and deployment of the intelligent model to reduce risk detection errors.

[0223] Specifically, existing methods typically rely on data from a small number of risk events to train intelligent models, limiting their coverage to samples of already occurred risks. This scenario proposes a deep representation learning scheme for drilling logging parameters that does not require manually labeled data. For logging time-series data (including key feature parameters such as total pool volume, riser pressure, and inlet flow rate), a sliding window dynamic standardization and engineering feature fusion technique is employed to automatically extract multi-dimensional feature vectors. This allows the intelligent model to provide early warnings of potential risks even when no risks have occurred, based on abnormal data fluctuations. Through an unsupervised learning framework, the distribution consistency of similar operating conditions is maximized within the feature space, while abnormal fluctuation patterns are separated. This overcomes the bottleneck of existing methods relying on a large number of labeled samples, making it particularly suitable for drilling scenarios where historical abnormal events are scarce.

[0224] Secondly, based on drilling engineering principles, directional correlation edges are constructed between parameters (e.g., key characteristic parameters). This aims to achieve more efficient graph network risk monitoring by establishing mechanism-driven parameter associations within the graph relationship. Causal relationship edges between key parameters can be established first based on fluid mass conservation and pressure balance equations. This mechanism-driven endogenous connection between parameters strengthens the relationship between parameters and reduces false alarms. The relationship between the difference between inlet and outlet flow rates driving the total pool volume, and the relationship between pump pressure fluctuations transmitted to casing pressure through the circulation system are established separately (resulting in a key characteristic parameter association graph). Furthermore, the theoretical values ​​and measured residuals of the physical equations can be calculated based on real-time data, and edge weights can be dynamically adjusted to achieve accurate and real-time graph update models. For example, during the pre-training phase, when the actual relationship between pump pressure and casing pressure deviates from the theoretical model, the edge weight is automatically reduced to weaken minor abnormal associations, ensuring the graph structure's sensitivity to drilling risks. The established drilling risk-sensitive parameter association graph will be used for the subsequent construction of a graph autoencoder.

[0225] Furthermore, a pre-training method for autoencoders (e.g., graph autoencoders) based on multi-dimensional time-series graphs is employed. This method includes: first, using a pre-constructed correlation graph of drilling risk-sensitive parameters as the prior structure of the GCN layer to achieve graph-constrained multi-dimensional time-series parameter linkage. Then, logging data nodes from consecutive time steps are connected sequentially according to the drilling process flow, forming a chain-like propagation path in the GCN layer to capture the gradual changes in parameters as drilling progresses. Next, a TCN layer is introduced, using dilated causal convolution to exponentially expand the receptive field, achieving multi-scale time-series perception. Afterward, the sequence, dimensionality-reduced by the GCN-TCN encoder, is decoded by the TCN-GCN decoder to obtain the decoded sequence mapping. Finally, the error between the original sequence and the reconstructed sequence is used as the loss function optimization objective to complete the pre-training of the graph autoencoder.

[0226] In addition, threshold rules for operating condition constraints (e.g., a preset set of operating condition thresholds) have been established, which may include the following steps: First, dynamic thresholds are formed by statistically analyzing the distribution characteristics of data in the training set and the current sliding window. Specifically, for the training set, the mean and standard deviation of all data are calculated, and the mean is multiplied by n times the standard deviation to form the "training set threshold" L. a Then, using a sliding window of length L, a reconstruction error of length L from the autoencoder output is selected. The mean and standard deviation of the data within the window are calculated, and the "real-time data threshold" G, which is the sum of the mean and n times the standard deviation, is taken. a Please refer to the following formula:

[0227]

[0228] Among them, L a The threshold for the training set, This is the statistical mean of the output set of the training set after the model has run. G is the statistical standard deviation of the training set after the model runs, where n is the standard deviation factor. a For the test set threshold, This is the statistical mean of the output set of the test set samples after the model runs. This is the statistical standard deviation of the test set samples in the output result set after the model runs.

[0229] By weighting and balancing the two thresholds with a certain proportional coefficient, the dynamic threshold can be obtained. Generally, the threshold of the training set is smaller, while the threshold of the test set is larger. By reasonably adjusting the weights of the two according to the physical constraints, a relatively suitable dynamic threshold calculation result can be achieved. The following formula (3) can be used as a reference:

[0230] T(t) = W x L a +W y G a (3)

[0231] Among them, W x For the threshold weights of the training set, W y The threshold weights are for the test set.

[0232] During drilling operation condition switching, rapid parameter changes often lead to high reconstruction errors in the autoencoder model output, causing the aforementioned dynamic threshold to become invalid for a short period, resulting in false alarms. Therefore, a condition-constrained dynamic threshold is introduced. First, on a dataset containing various normal operating conditions, an operating condition recognition algorithm is used to label the current operating condition. Then, the dataset classified by operating condition is input into the autoencoder to obtain the reconstruction errors for each operating condition, and the "training set threshold" L under this operating condition is statistically determined. a .

[0233] Furthermore, a fuzzy assessment method for geological-engineering pre-drilling assessment is introduced, including: First, determining a type of drilling risk to be monitored (e.g., blowout) and a set of risk event elements. The set of risk event elements includes the geological and engineering factors that need to be considered in this assessment. Geological factors generally include formation lithology, formation pressure coefficient, geological stability, etc., while engineering factors include the relationship between design ECD and formation pressure, the severity of risks occurring in adjacent wells, etc. Then, a risk rating scoring table for the set of risk event elements is established, which is divided into at least three categories: "low risk", "medium risk", and "high risk", and threshold ranges and risk levels are defined, and a comparison judgment matrix is ​​established. For information that has been interpolated to this well, the risk score can be obtained directly by classification; for the risk information of adjacent wells, the degree of influence on this well needs to be calculated using a distance-weighted method (see formula (1) and formula (2)), so as to calculate the risk item of this well according to the following formula:

[0234]

[0235] Among them, w i For distance weights, l i r is the distance between the adjacent well and the target well. i For the quantitative risk score of adjacent wells, r a This is the final risk score for the risk factor.

[0236] By solving the comparison judgment matrix at this depth, the risk assessment score R(H) for the current depth can be obtained. This allows the construction of a geological-engineering pre-drilling assessment profile (e.g., a pre-defined geological-engineering pre-drilling assessment profile) for the target wellbore's designed wellbore trajectory.

[0237] Finally, in practical implementation, when the real-time data stream is input into the algorithm, the algorithm will generate a dynamic threshold in real time. First, referring to formulas (1), (2), and (3), the parameters of the dynamic threshold are constructed: training set threshold L a Training set threshold G a .

[0238] When the time point is t=n, the drill bit is located at depth H. A geological-engineering pre-drilling assessment profile can be established to read the corresponding risk assessment score R(H), which is then used as a primary dependency in the dynamic threshold formula (Equation 6) to establish the dynamic threshold for geological-engineering risk embedding: T n =T(X) n ) = W y L a +W z (G a )+L a (1-R(H))(6).

[0239] Then, by comparing the dynamic threshold of the geological-engineering risk embedding with the reconstruction error output of the autoencoder model, it is determined whether there is a drilling risk, so as to decide whether to issue an alarm.

[0240] Based on the above ideas, in specific implementation, please refer to the following overall guidelines. Figure 10 As shown, the following steps may be included:

[0241] The first step is to combine correlation analysis with actual engineering constraints to select key feature parameters as inputs and determine the physical constraint relationships between parameters; at the same time, select datasets under normal working conditions that are relatively easy to retrieve in engineering and complete data preprocessing.

[0242] The second step is to establish a drilling risk-sensitive parameter graph structure based on physical constraints (e.g., a correlation graph of key characteristic parameters) to constrain the correlation between multidimensional parameters.

[0243] The third step is to learn and model the temporal characteristics of key engineering logging parameters under various normal working conditions by pre-training the GCN-TCN graph autoencoder (to obtain an intermediate detection model).

[0244] The fourth step involves connecting the trained graph autoencoder to real-time engineering logging data streams, selecting data within a specified time window for inference, and returning the reconstruction error. This error is then optimized using a dynamic threshold derived from the geological-engineering risk embedding method to achieve risk monitoring. The risk monitoring results can be found in [reference needed]. Figure 12 As shown in the figure, the GCN-TCN-AE (or GCN-TCN) alarm represents the monitoring result when drilling risk monitoring is performed solely using the preset detection model with a graphical autoencoder provided in this manual. The dynamic threshold GCN-TCN-AE alarm represents the monitoring result when drilling risk monitoring is performed using both the preset detection model with a graphical autoencoder provided in this manual and the dynamic threshold.

[0245] Figure 10 The "dynamic threshold determination" is one of the core innovations of this solution. (See attached document.) Figure 11 As shown, it can specifically include the following:

[0246] S1: Establish dynamic thresholds for operating condition constraints (e.g., a preset set of operating condition thresholds). To address the potential differences in the degree of abnormality under different operating conditions and the limited fluctuations in reconstruction errors caused by operating condition switching, training set thresholds for different operating conditions are pre-determined through operating condition identification.

[0247] S2: Establish a geological-engineering pre-drilling assessment profile (e.g., a pre-defined geological-engineering pre-drilling assessment profile). Establish a set of risk event elements for each detected risk type, and use a fuzzy matrix method to create the geological-engineering pre-drilling assessment profile.

[0248] S3: Constructing Dynamic Thresholds. Based on the condition-based training set thresholds established in Step 1, statistical analysis is performed on the selected data from the training set and real-time data stream to establish training set thresholds and test set thresholds. Combining the geological-engineering pre-drilling assessment profile established in Step 2, the dynamic threshold at the current moment (e.g., the target dynamic threshold) is calculated.

[0249] See Figure 10 As shown, the training and testing of graph autoencoders can be divided into the following contents.

[0250] S1: Select the best input features and perform data preprocessing on the logging data to form a data processing flow that includes outlier detection, interpolation, and smoothing filtering.

[0251] Before constructing a risk diagnosis model, the primary task is to select optimal input features and perform data preprocessing. This embodiment extensively collected and organized comprehensive well logging time-series data of over 120 blowout events that occurred during drilling operations in more than 100 wells across five major oil drilling areas in China, as well as well logging data under normal operating conditions. This data covers detailed records of various fluids, including oil, gas, and water, invading the formation, providing valuable information resources for in-depth analysis of blowout events.

[0252] Preliminary data analysis revealed no single characteristic significantly correlated with the occurrence of overflow, such as... Figure 11 The correlation calculation results are shown. Therefore, to ensure that the model can be quickly and effectively applied in the drilling field, the following key features are selected as inputs for model training: riser pressure, inlet flow rate, outlet flow rate, total pool volume, casing pressure, and total hydrocarbon content.

[0253] Standby pressure is considered because pressure changes are a crucial indicator for monitoring downhole dynamics during drilling operations. When a blow occurs, the drilling fluid begins to flow freely, disrupting the equilibrium of the hydrostatic column. As the drilling fluid rises, the hydrostatic column height decreases, leading to a drop in bottomhole pressure. However, inside the drill string, the impact of fluid intrusion is minimal, and the hydrostatic pressure and circulation pressure loss remain almost constant. Therefore, the decrease in bottomhole pressure results in a decrease in standby pressure, a change that is significant for predicting blows.

[0254] Changes in inlet and outlet flow rates can also reflect the occurrence of overflow risk. Under normal circumstances, inlet and outlet flow rates should remain relatively stable. However, when an overflow occurs, the outlet flow rate will increase without an increase in the inlet flow rate, which is an important characteristic for identifying an overflow.

[0255] Total pool volume, as a key parameter describing drilling fluid reserves, also significantly impacts blowout prediction. As drilling fluid is returned, the total pool volume gradually increases. In the case of oil-water intrusion, the total pool volume increases accordingly with the increase in outlet flow rate; similarly, in the case of gas intrusion, the total pool volume increases due to the expansion of gas bubbles and the resulting increase in outlet flow rate. These changes become apparent shortly after oil-water intrusion and gas intrusion, providing crucial clues for blowout prediction.

[0256] After selecting these key features, a data cleaning step was performed to reduce the impact of noise, outliers, and missing values. Data cleaning is a crucial step in machine learning preprocessing, improving data quality and making it more suitable for subsequent analysis and modeling. In this embodiment, outlier detection, missing value handling, data smoothing, and normalization were implemented through code, laying a solid foundation for subsequent model training.

[0257] In anomaly detection tasks, the scarcity of anomalous data is a common problem. When normal data far outweighs anomalous data, supervised models may become lazy, tending to predict all data as normal. This model avoids this problem through unsupervised learning, ensuring that a large amount of data from normal drilling periods is fully utilized. In this embodiment, drilling logs are used to distinguish between normal wells or normal operating conditions, as well as risk periods. Normal data is used to train the unsupervised model, forming multiple anomalous datasets for each overflow event. Each risk event dataset is divided into "normal period - risk period" datasets at a ratio of 40% and 60%, respectively, to achieve the purpose of detecting the quality of risk monitoring.

[0258] To verify the model's accuracy across different wells, we adopted a cross-well data partitioning strategy, ensuring that the training and test sets use data from different wells. This improves the model's generalization ability, enabling it to better adapt to unseen well conditions.

[0259] S2: Establish an overflow-sensitive parameter graph structure based on physical constraints (e.g., a key feature parameter association graph based on physical constraints).

[0260] Subsequently, based on the selected parameters, this example establishes an overflow sensitivity parameter graph structure based on physical constraints, as shown below. Figure 6 As shown, several sensitive engineering logging parameters were constructed using physical constraints to establish graph edge weight relationships. Physical formulas and specific working condition constraints were embedded, and the base layer was frozen for GCN network construction. By establishing the graph structure relationship of sensitive parameters with physical constraints, heterogeneous errors caused by switching between normal working conditions and fluctuations within working conditions were filtered out, enhancing model accuracy.

[0261] S3: Establish an intelligent overflow risk identification model based on graph autoencoders.

[0262] After completing the construction of the training set, risk test set, and sensitive parameter graph structure, we entered the model building and testing phase, constructing a graph convolutional autoencoder, the structure of which is as follows: Figure 2 As shown in the figure. In this embodiment, GCN (Graph Convolutional Network Unit) and TCN (Temporal Convolutional Network Unit) are selected as the base layers of the encoder and decoder of the unsupervised temporal model. GCN performs well in handling parameter associations and constraints, while TCN has a significant advantage in handling long-term relationships in time series and can capture temporal dependencies in the data.

[0263] S4: Use the dynamically threshold-optimized model for actual testing.

[0264] Taking the normal operating condition dataset consisting of 32 wells in this example, the original sequence of the normal operating condition data is input into an autoencoder. After dimensionality reduction by a decoder, it is reconstructed by the encoder to form a reconstructed sequence. The error between the reconstructed sequence and the original sequence is called the reconstruction error. The autoencoder aims to minimize the reconstruction error to achieve a smaller reconstruction error output on the representation of the normal sequence. In this way, when an abnormal sequence is input, the model will output a larger reconstruction error, thereby achieving risk monitoring.

[0265] Combination Figure 11 The main methods for constructing the aforementioned dynamic thresholds include the following:

[0266] S1: Establish dynamic thresholds for operating condition constraints.

[0267] Taking the normal operating condition dataset consisting of 32 wells in this example as an example, an operating condition identification algorithm is introduced to separate various drilling operating conditions (such as drilling, circulation, pump start-up, and pump stop-up), and to count the reconstruction error of the data in the autoencoder under various operating conditions, and to count it as the training set threshold for that operating condition.

[0268] S2: Establish a geological-engineering pre-drilling assessment profile based on fuzzy matrices.

[0269] This example establishes a dynamic threshold combining geological and engineering risks. First, a geological-engineering pre-drilling assessment profile considering multiple drilling risk factors must be created. In this example, four factors are selected as assessment objects: adjacent well risk coefficient, formation pressure coefficient, formation lithology, and the difference between the design ECD and formation pressure. Risk level scores are assigned to these factors, and a comparative judgment matrix for risk assessment is formed. By solving this matrix, a geological-engineering pre-drilling assessment profile considering multiple geological and engineering risks can be generated.

[0270] S3: Establish a dynamic threshold for embedding geological-engineering information.

[0271] For the trained GCN-TCN autoencoder risk monitoring model (hereinafter referred to as the "risk monitoring model"), this example selects a dataset of 12 wells used for training and inputs it into the model to obtain a set of reconstruction errors. This example uses a sliding window to calculate the real-time test set threshold. Specifically, a window of length L is selected, and the data points from the most recent 60 time steps are statistically analyzed, with a standard deviation coefficient of n=1. Simultaneously, the data stream enters the working condition identification algorithm to identify the current working condition. Based on the working condition, the currently calculated training set threshold is determined, with a standard deviation coefficient of n=1. The optimal weight value used in the test is W. y =0.3, W z=0.7. The current vertical depth of the drill bit is H, which is calculated by the logging tool. The risk value R(H) of the current vertical depth can be read from the constructed geological-engineering pre-drilling assessment profile R. Substitute this into formula (6) to obtain the exact value of the current dynamic threshold.

[0272] By introducing a dynamic threshold embedded in the pre-drilling assessment of geological and engineering factors, when the reconstruction error fluctuates slightly due to changes in working conditions or parameter fluctuations, the dynamic threshold will dynamically control the alarm limit according to the characteristics of the working conditions and the constraints of the pre-drilling assessment. In low-lying formations, the dynamic threshold will be more lenient overall to reduce the false alarm rate; while in dangerous formations it will be more stringent to ensure timely alarms. Ultimately, a dynamic threshold embedded with geological and engineering information was established.

[0273] Through the above scenario examples, the drilling risk determination method provided in this specification has been verified to have the following advantages over existing methods: 1. Reduced sample requirements: This method adopts an unsupervised training approach, avoiding reliance on a small amount of risk data and making full use of a large amount of normal drilling data to train an unsupervised time series model. 2. Graph embedding of drilling risk sensitive parameters: By introducing explicit physical constraints, a correlation graph of drilling risk sensitive parameters is established to strengthen the correlation constraints between parameters. 3. Unsupervised monitoring using a multidimensional time series graph autoencoder: This method uses a graph convolutional network to capture short-term relationships between sequences and connects time-domain convolutions to realize medium- and long-term time series relationships, capturing the medium- and long-term characteristics of drilling risk accumulation. 4. Fuzzy assessment method for geological-engineering risks: By introducing a fuzzy assessment method, multi-source heterogeneous risk factors are structured to form a geological-engineering pre-drilling assessment profile for the target well. 5. Dynamic threshold optimization for geological-engineering risk embedding: By introducing the geological-engineering pre-drilling assessment profile, dynamic thresholds and macro-geological and adjacent well information are realized, improving the monitoring accuracy of high-risk strata and reducing false alarms caused by fluctuations due to operating condition switching and parameter fluctuations. Therefore, it not only improves the prediction accuracy of the model, but also provides a strong guarantee for the safety and efficiency of drilling operations.

[0274] While this specification provides the steps of operation for the methods described in the embodiments or flowcharts, more or fewer steps may be included based on conventional or non-inventive means. The order of steps listed in the embodiments is merely one possible order of execution among many steps and does not represent the only possible order. In actual device or client product execution, the methods shown in the embodiments or drawings may be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment). The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitations, the presence of other identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded. The terms "first," "second," etc., are used to denote names and do not indicate any particular order.

[0275] Those skilled in the art will also know that, besides implementing the controller using purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller function as logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices within it used to implement various functions can also be considered structures within that hardware component. Alternatively, the devices used to implement various functions can be considered as both software modules implementing the method and structures within a hardware component.

[0276] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer-readable storage media, including storage devices.

[0277] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this specification can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of this specification can essentially be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments of this specification.

[0278] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. This specification can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0279] Although this specification has been described by way of examples, those skilled in the art will recognize that many variations and modifications are possible without departing from the spirit of this specification, and it is intended that the appended claims cover such variations and modifications without departing from the spirit of this specification.

Claims

1. A method of determining a drilling risk, characterized by, The method comprises the following steps: obtaining target well current target logging data and a current time period logging data set; determining the current working condition of the target well according to the target logging data; determining the target well current target dynamic threshold according to the current working condition of the target well, the current time period logging data set, and a preset geology-engineering pre-drilling evaluation profile; the preset geology-engineering pre-drilling evaluation profile is a template profile capable of estimating the evaluation score of the corresponding drilling risk based on the depth domain, which is obtained by jointly considering the influence of the associated risk accident factors from the macro dimension and the mutual influence between the well and the adjacent well through fuzzy evaluation; obtaining the corresponding target reconstruction error by processing the target logging data by using the preset detection model; wherein the preset detection model is a model trained by unsupervised learning based on a graph autoencoder; the graph autoencoder at least comprises an encoder, and the encoder at least comprises a GCN layer and a TCN layer connected in sequence; determining whether the target well has a drilling risk according to the target reconstruction error and the target dynamic threshold; wherein, determining the target well current target dynamic threshold according to the current working condition of the target well, the current time period logging data set, and the preset geology-engineering pre-drilling evaluation profile comprises: determining the matching target training set threshold according to the current working condition of the target well and querying the preset working condition threshold set; determining the corresponding target test set threshold by processing the current time period logging data set by using the graph autoencoder in the preset detection model; determining the current drill bit vertical depth in the target well according to the target logging data; and determining the corresponding target risk evaluation value according to the current drill bit vertical depth and the preset geology-engineering pre-drilling evaluation profile; determining the target well current target dynamic threshold according to the target training set threshold, the target test set threshold, and the target risk evaluation value; the target test set threshold is determined in the following manner: converting the current time period logging data set into a corresponding data stream; extracting logging data at a plurality of continuous time points from the data stream by using a sliding time window, and processing the logging data at the plurality of continuous time points by using the graph autoencoder in the preset detection model to obtain a corresponding processing result; and determining a test set threshold capable of representing the current data fluctuation trend and having strong real-time performance for the current time point according to the processing result, as the target test set threshold.

2. The method of claim 1, wherein, The method further comprises: obtaining sample logging data; and constructing a sample training set according to the sample logging data; determining key feature parameters for drilling risk by correlation analysis according to the sample logging data; determining the association relationship of the key feature parameters with drilling according to drilling engineering rules and fluid mass conservation and pressure balance rules; and constructing a key feature parameter association graph for drilling risk based on physical constraints according to the association relationship; constructing an initial detection model based on the graph autoencoder; performing prior learning on the initial detection model by using the key feature parameter association graph, so that the GCN layer learns the association constraints of different key feature parameters to obtain an intermediate detection model; The intermediate detection model is unsupervised learned by using the sample training set, and a preset detection model meeting a requirement is obtained.

3. The method of claim 2, wherein, The key feature parameters include at least one of standpipe pressure, inlet flow, outlet flow, total pool volume, and casing pressure.

4. The method of claim 2, wherein, The graph autoencoder further comprises a decoder; wherein the decoder is connected with the encoder, and the decoder at least comprises a TCN layer and a GCN layer connected in sequence.

5. The method of claim 2, wherein, According to sample logging data, a sample training set is constructed, including: First type sample logging data corresponding to a risk working condition and second type sample logging data corresponding to a normal working condition are screened from the sample logging data; According to the second type sample logging data, a sample training set is constructed; wherein the second type sample logging data is used to construct a test data set.

6. The method of claim 1, wherein, The method further comprises: Target geological background data of a region where the target well is located and target historical drilling records of a drilled adjacent well adjacent to the target well are obtained; According to the target geological background data and the target historical drilling records, a risk accident factor set associated with drilling risks is determined; wherein the risk accident factor set at least includes geological factors and engineering factors; According to the risk accident factor set, a risk grade fuzzy evaluation rule is determined; According to the risk grade fuzzy evaluation rule, a comparison judgment matrix of risk evaluation is generated; According to the comparison judgment matrix of risk evaluation, a preset geological-engineering pre-drilling evaluation profile is constructed.

7. A method of determining a drilling risk, characterized by, It comprises: Target logging data of the target well at present and a logging data group in a current time period are obtained; According to the target logging data, a current working condition of the target well is determined; According to the current working condition of the target well, the logging data group in the current time period, and the preset geological-engineering pre-drilling evaluation profile, a target dynamic threshold value of the target well at present is determined; the preset geological-engineering pre-drilling evaluation profile is a template profile capable of estimating an evaluation score of corresponding drilling risks based on a depth domain, which is constructed by considering the influence of associated risk accident factors from a macroscopic dimension and the mutual influence between the well and adjacent wells through fuzzy evaluation; According to the target logging data, a corresponding target result data is obtained by processing the target logging data by using the preset detection model; wherein the preset detection model at least contains a graph autoencoder; According to the target result data and the target dynamic threshold value, it is determined whether the target well has drilling risks; According to the current working condition of the target well, the logging data group in the current time period, and the preset geological-engineering pre-drilling evaluation profile, a target dynamic threshold value of the target well at present is determined, including: according to the current working condition of the target well, a target training set threshold value matched is determined by querying a preset working condition threshold set; a target test set threshold value corresponding is determined by processing the logging data group in the current time period by using the graph autoencoder in the preset detection model; a current drill bit vertical depth in the target well is determined according to the target logging data; and a corresponding target risk evaluation value is determined according to the current drill bit vertical depth and the preset geological-engineering pre-drilling evaluation profile; and the target dynamic threshold value of the target well at present is determined according to the target training set threshold value, the target test set threshold value, and the target risk evaluation value. The target test set threshold is determined in the following manner: converting the logging data set of the current time period into a corresponding data stream; using a sliding time window to extract logging data of multiple continuous time points from the data stream, and processing the logging data of the multiple continuous time points by using a graph autoencoder in a preset detection model to obtain a corresponding processing result; and according to the processing result, determining a test set threshold for the current time point that can represent a current data fluctuation trend and has strong real-time performance as the target test set threshold.

8. An apparatus for determining a drilling risk, characterized by Comprise: An acquisition module configured to acquire target logging data of a target well at a current time point and a logging data set of a current time period; A first determination module configured to determine a current working condition of the target well according to the target logging data; A second determination module configured to determine a target dynamic threshold of the target well at the current time point according to the current working condition of the target well, the logging data set of the current time period, and a preset geology-engineering pre-drilling evaluation profile; the preset geology-engineering pre-drilling evaluation profile is a template profile that can estimate an evaluation score of a corresponding drilling risk based on a depth domain, and is obtained by considering the influence of associated risk accident factors from a macro dimension and the mutual influence between the well and adjacent wells through fuzzy evaluation; A processing module configured to obtain a corresponding target reconstruction error by processing the target logging data by using a preset detection model; the preset detection model is a model trained by unsupervised learning based on a graph autoencoder; the graph autoencoder at least includes an encoder, and the encoder at least includes a GCN layer and a TCN layer connected in sequence; A third determination module configured to determine whether the target well has a drilling risk according to the target reconstruction error and the target dynamic threshold; The second determination module is specifically configured to: query a preset working condition threshold set according to the current working condition of the target well to determine a matching target training set threshold; determine a corresponding target test set threshold by processing the logging data set of the current time period by using the graph autoencoder in the preset detection model; determine a current bit vertical depth in the target well according to the target logging data; and determine a corresponding target risk evaluation value according to the current bit vertical depth and the preset geology-engineering pre-drilling evaluation profile; and determine the target dynamic threshold of the target well at the current time point according to the target training set threshold, the target test set threshold, and the target risk evaluation value; The target test set threshold is determined in the following manner: converting the logging data set of the current time period into a corresponding data stream; using a sliding time window to extract logging data of multiple continuous time points from the data stream, and processing the logging data of the multiple continuous time points by using a graph autoencoder in a preset detection model to obtain a corresponding processing result; and according to the processing result, determining a test set threshold for the current time point that can represent a current data fluctuation trend and has strong real-time performance as the target test set threshold.

9. A computer-readable storage medium, characterized in that, A computer readable storage medium having stored thereon computer instructions, which when executed by a processor, implement the steps of the method of any one of claims 1 to 7.

Images