Drilling risk identification model, establishment method, identification method and computer device

Abstract

The application provides a drilling risk identification model and an establishment method, an identification method and a computer device. The drilling risk identification model establishment method comprises the following steps: establishing a database containing risk characterization parameters; establishing a composite network model based on the database, wherein the model comprises a convolution network layer and an RBF network layer; and training the network model to obtain a drilling risk identification model. The model is obtained by the above method. The identification method is realized based on the above model. The device comprises a processor and a memory storing a computer program, and when the processor executes, the above establishment method is realized, or the above identification method is realized. The application can quickly and effectively identify the gas drilling risk, which is beneficial to maintaining the normal operation of drilling and reducing the drilling cost.

Description

Technical Field

[0001] This invention relates to the field of oil and gas drilling, and more specifically, to a drilling risk intelligent identification model, a method for establishing a drilling risk intelligent identification model, a drilling risk intelligent identification method, and a computer device and computer-readable storage medium for implementing the drilling risk intelligent identification method. Background Technology

[0002] In recent years, drilling technology has developed rapidly in China, but some safety issues have also emerged. Currently, the means of monitoring safety risks in drilling technology remain insufficient. When monitoring data shows anomalies, monitoring personnel need to make timely and accurate judgments. This leads to a strong degree of subjectivity and lag in the on-site diagnosis of abnormal risks. Furthermore, the differences in drilling conditions across different regions result in significant variations in the characteristics of relevant parameters reflecting the occurrence of abnormal risks, further increasing the difficulty of on-site monitoring. How to identify and address drilling safety risks in a timely and accurate manner has become a crucial aspect of ensuring the success rate of gas drilling.

[0003] Currently, research on intelligent identification of drilling risks for safety monitoring is lacking. For example, a patent document published on December 1, 2020, entitled "A Method for Early Warning of Drilling Spills in Oil and Gas Wells Based on Long Short-Term Memory Networks" (publication number CN112016766A), describes a method for early warning of drilling spills in oil and gas wells based on long short-term memory networks. This method includes the following steps: constructing a spill risk early warning model based on long short-term memory networks, which includes an input layer, a hidden layer, and an output layer; training the spill risk early warning model using existing drilling risk data; and using the trained spill risk early warning model to provide early warning of drilling spills in oil and gas wells. This method can reduce the reliance on prior knowledge and expert experience for risk identification; however, it is mainly applicable to early warning of drilling spill risks and has poor applicability to other drilling risks.

[0004] A patent document published on November 12, 2019, entitled "Drilling Leakage Risk Identification Method, System, and Equipment Based on Convolutional Neural Networks" (publication number CN110443488A), describes a drilling leakage risk identification method based on convolutional neural networks. This method acquires real-time operating condition data collected by a comprehensive logging instrument; expands the dataset for each real-time operating condition data acquired by the comprehensive logging instrument using a sliding overlap sampling method; performs normalization preprocessing on the real-time operating condition data; and inputs the normalized preprocessed real-time operating condition data into a trained convolutional neural network model, finally outputting the drilling leakage risk identification result. Although this method utilizes a convolutional neural network model, its application of the model is relatively superficial. The initial cluster center design of the convolutional neural network is unreasonable, with excessively wide center widths, which easily leads to model training failure, causing identification errors and hindering fast and effective identification. Furthermore, its application scope is narrow, with a low identification rate for other downhole risks. Both of the above methods use new information technologies to reduce human factors in the risk identification process, but their application scope is narrow, they have not improved the convolutional neural network model, have low recognition rates, and are prone to recognition failures. Summary of the Invention

[0005] The purpose of this invention is to address at least one of the aforementioned shortcomings of the prior art. For example, this invention provides a method for establishing a drilling risk identification model to overcome the defects of slow drilling risk identification and inaccurate identification in the prior art.

[0006] To achieve the above objectives, the present invention provides a method for establishing a drilling risk identification model, the method comprising the following steps: establishing a database containing risk characterization parameters; establishing a composite network model based on the database, the model comprising a convolutional neural network layer and an RBF neural network layer; wherein the convolutional neural network layer comprises a first convolutional layer and a second convolutional layer, the first convolutional layer being able to extract the feature information implicit in each parameter itself, the second convolutional layer being able to extract the implicit relationship between each parameter, and being able to combine feature information of different scales to obtain composite feature information; the RBF neural network layer being able to classify the composite feature information; and training the composite network model to obtain a drilling risk identification model.

[0007] In an exemplary embodiment of the drilling risk identification model establishment method of the present invention, both the first and second convolutional layers use three convolutional kernels of different scales.

[0008] In an exemplary embodiment of the drilling risk identification model establishment method of the present invention, after establishing the composite network model, the method further includes a step of optimizing the model.

[0009] In an exemplary embodiment of the method for establishing a drilling risk identification model according to the present invention, the optimization may include the optimization method of the K-means algorithm.

[0010] In an exemplary embodiment of the drilling risk identification model establishment method of the present invention, the training includes: training and updating the cluster centers of the K-means algorithm based on the loss function components of the distance mean square error.

[0011] In an exemplary embodiment of the drilling risk identification model establishment method of the present invention, a criterion for selecting the number of hidden layer nodes is established during the training to improve the computational efficiency and generalization of the model.

[0012] In another aspect, the present invention provides a drilling risk identification model, which can be obtained by the drilling risk identification model establishment method described above.

[0013] In another aspect, the present invention provides a method for intelligent identification of drilling risks. The method may include the following steps: acquiring multiple risk characterization parameters; extracting feature information of each parameter and the relationship between each parameter using a convolutional neural network to obtain composite feature information; classifying the composite feature information using an RBF neural network; and performing real-time risk identification of drilling risks by using the relationship between the number of hidden layer nodes and evaluation values ​​for different risks to obtain the probability of risk occurrence of the risk characterization parameter model.

[0014] In another aspect, the present invention provides a method for intelligent identification of drilling risks, wherein the RBF neural network is an optimized and trained network; wherein the optimization may include: initializing the cluster centers and center widths of the RBF neural network using the K-means algorithm; the training may include: training and updating the cluster centers of the K-means algorithm based on the loss function components of the mean squared distance error.

[0015] In an exemplary embodiment of the drilling risk intelligent identification method of the present invention, a selection criterion for the number of hidden layer nodes is established in the intelligent identification method to improve the computational efficiency and generalization of the model.

[0016] In an exemplary embodiment of the drilling risk intelligent identification method of the present invention, the hidden layer selection rule may include: assigning weights to the accuracy of network model training and the training time of network model, and determining whether the selection of hidden layer nodes is reasonable based on the ratio of the two parameters corresponding to different numbers of hidden layer nodes.

[0017] In an exemplary embodiment of the drilling risk intelligent identification method of the present invention, the selection rule for the number of hidden layer nodes may include: (1) determining the number of input layers of the convolutional network and the number of output layers of the RBF neural network, and determining the basic network model, which includes a convolutional network and an RBF neural network; (2) according to Determine the initial number of hidden layer nodes in the RBF neural network, where l is the number of input nodes in the RBF network; r is the number of hidden layer nodes; k is the number of output layer nodes in the RBF network; ε is a constant coefficient with a value range of 0-10; (3) Train the basic network model to obtain the final training time T and the accuracy B of the basic network model test set. In order to eliminate the influence of the dimensions between the indicators, the training time and the accuracy B of the network model test set are normalized to 0-1; (4) Use the following formula to obtain the evaluation value C(r) corresponding to different numbers of hidden layer nodes, where, μ and ν are constant coefficients; (5) For different risk types, set different increments of the number of hidden layer nodes Δr, and repeat step (4) to obtain a new evaluation value C(r+1). If the new evaluation value continues to decrease or tends to remain unchanged, stop increasing the number of hidden layer nodes; (6) Normalize the evaluation values ​​corresponding to different numbers of hidden layer nodes by 0-1, and take the number of hidden layer nodes corresponding to the point where the normalized evaluation value C(r) is 100% as the optimal number of nodes.

[0018] In an exemplary embodiment of the drilling risk intelligent identification method of the present invention, the acquisition of multiple risk characterization parameters includes safety risk type and sampling time. Specifically, when the risk type is formation gas production, the multiple risk characterization parameters may include: pipeline pressure, methane concentration, pipeline differential pressure, and oxygen concentration; when the risk type is formation water production, the multiple risk characterization parameters may include hook load, riser pressure, and outlet humidity; and when the risk type is stuck drill bit risk, the multiple risk characterization parameters may include rotary table speed and rotary table torque.

[0019] In an exemplary embodiment of the drilling risk intelligent identification method of the present invention, the convolutional neural network may include a first convolutional network layer and a second convolutional network layer, wherein the first convolutional network layer is used to extract the feature information implicit in each parameter of the parameter model itself; and the second convolutional network layer is used to extract the implicit relationship between each parameter of the parameter model.

[0020] In an exemplary embodiment of the drilling risk intelligent identification method of the present invention, the kernel size of the first convolutional network layer C1 can be m*1, that is, the time feature information of a certain feature parameter in m rows and 1 columns, where m is the vertical length of the convolutional kernel; the kernel size of the second convolutional network layer C2 can be 1*n, used to extract the correlation feature information between parameters of a certain risk in the characterization parameter in 1 row and n columns, where n is the number of various safety risk feature parameters.

[0021] In another aspect, the present invention provides a computer device, which may include: a processor; a memory storing a computer program, wherein when the computer program is executed by the processor, it implements the drilling risk identification model establishment method as described above, or the drilling risk intelligent identification method as described above.

[0022] In another aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method for establishing a drilling risk identification model as described above, or the method for intelligent identification of drilling risks as described above.

[0023] Compared with the prior art, the beneficial effects of the present invention may include at least one of the following:

[0024] (1) This invention uses an algorithm optimization method to initialize the cluster centers and center width of the RBF network layer, which solves the problem that the model training cannot converge;

[0025] (2) The present invention can quickly and effectively identify risks that occur during drilling, thereby reducing drilling costs;

[0026] (3) This invention improves the efficiency of drilling risk identification, reduces manual intervention, improves detection accuracy, and can accelerate drilling accuracy and save work cycle. Attached Figure Description

[0027] The above and other objects and / or features of the present invention will become clearer from the following description taken in conjunction with the accompanying drawings, in which:

[0028] Figure 1 A flowchart illustrating an exemplary embodiment of the drilling risk intelligent identification method of the present invention is shown.

[0029] Figure 2 The diagram shows a two-dimensional spatial projection of data points in an exemplary embodiment of the intelligent identification method for drilling risks according to the present invention.

[0030] Figure 3 A two-dimensional Gaussian kernel function graph is shown as an exemplary embodiment of the intelligent identification method for drilling risks according to the present invention.

[0031] Figure 4 A neural network model structure diagram of an exemplary embodiment of the drilling risk intelligent identification method of the present invention is shown.

[0032] Figure 5A This diagram illustrates a formation gas production model showing the relationship between the number of hidden layer nodes and evaluation values ​​for different risks, representing an exemplary embodiment of the intelligent drilling risk identification method of the present invention.

[0033] Figure 5B The diagram shows a water discharge model of the relationship between the number of hidden layer nodes and the evaluation value for different risks in an exemplary embodiment of the drilling risk intelligent identification method of the present invention.

[0034] Figure 5C The diagram illustrates a stuck drill model showing the relationship between the number of hidden layer nodes and evaluation values ​​for different risks in an exemplary embodiment of the intelligent drilling risk identification method of the present invention.

[0035] Figure 6A The diagram shows the loss values ​​of the training results of the formation gas production risk training set in an exemplary embodiment of the drilling risk intelligent identification method of the present invention.

[0036] Figure 6B The diagram shows the accuracy of the training results of the formation gas production risk training set in an exemplary embodiment of the drilling risk intelligent identification method of the present invention.

[0037] Figure 7A The diagram shows the loss values ​​of the training results of the formation water production risk training set in an exemplary embodiment of the drilling risk intelligent identification method of the present invention.

[0038] Figure 7B The diagram shows the accuracy of the formation water production risk training set training results for an exemplary embodiment of the drilling risk intelligent identification method of the present invention.

[0039] Figure 8A The diagram shows the loss values ​​of the training results of the stuck pipe risk training set in an exemplary embodiment of the drilling risk intelligent identification method of the present invention.

[0040] Figure 8B The diagram shows the accuracy of the training results of the stuck pipe risk training set in an exemplary embodiment of the intelligent identification method for drilling risks according to the present invention.

[0041] Figure 9 A schematic diagram of a computer device structure is shown as an exemplary embodiment of a drilling risk intelligent identification device according to the present invention.

[0042] Explanation of reference numerals in the attached figures:

[0043] 100 - Computer equipment; 110 - Memory; 120 - Processor. Detailed Implementation

[0044] The following will describe in detail, with reference to the accompanying drawings and specific embodiments, a drilling risk identification model and its establishment method, identification method and computer equipment of the present invention.

[0045] It should be noted that "first," "second," etc., are merely for the convenience of description and distinction, and should not be interpreted as indicating or implying relative importance.

[0046] Exemplary Example 1

[0047] This exemplary embodiment provides a method for establishing a drilling risk identification model.

[0048] The method for establishing the drilling risk identification model may include the following steps:

[0049] Establish a database containing risk characterization parameters.

[0050] Based on the database, a composite network model is established, which includes convolutional neural network layers and RBF neural network layers.

[0051] The convolutional neural network layer includes a first convolutional layer and a second convolutional layer. The first convolutional layer can extract the feature information implied by each parameter itself, and the second convolutional layer can extract the relationship implied between each parameter and combine feature information of different scales to obtain composite feature information.

[0052] In this embodiment, both the first and second convolutional layers use three convolutional kernels of different sizes.

[0053] The RBF neural network layer can classify the composite feature information.

[0054] The composite network model is trained to obtain the drilling risk identification model.

[0055] In this embodiment, a selection criterion for the number of hidden layer nodes is established during training to improve the computational efficiency and generalization ability of the model. This selection criterion is pre-defined in the algorithm before execution and is not included in the RBF neural network layers; it can be run independently.

[0056] The selection criterion for the number of hidden layer nodes can be applied to the entire risk identification model. However, since only the RBF neural network layer in this model has hidden layer nodes, it can also be understood as acting on the RBF neural network layer.

[0057] The hidden layer selection rule proposed in this embodiment can assign weights to the test set accuracy and training time of the network model, and determine whether the selection of hidden layer nodes is reasonable based on the ratio of the two parameters corresponding to different numbers of hidden layer nodes.

[0058] By training the hidden layer nodes of the RBF neural network using the hidden layer selection rule proposed in this embodiment, the relationship between the evaluation value of different risks and the number of hidden layer nodes can be obtained. Those skilled in the art can, based on experience, select the increment of the number of hidden layer nodes for different risk models to optimize the identification and evaluation results of the drilling risk identification model.

[0059] In this embodiment, after establishing the composite network model, the above method also includes an optimization step for the model.

[0060] In this embodiment, the model can be optimized using the K-means algorithm, and then the cluster centers of the K-means algorithm can be trained and updated based on the loss function component of the distance mean square error.

[0061] Furthermore, the optimization of the K-means algorithm can be the same as that in Exemplary Example 3.

[0062] Furthermore, the training of the loss function components based on the mean squared error of the distance can be the same as in Exemplary Example 3.

[0063] In this embodiment, the risk characterization parameters can be the same as those in Exemplary Embodiment 3.

[0064] In this embodiment, the first and second convolutional layers can be the same as those in Exemplary Embodiment 3.

[0065] In this embodiment, the RBF neural network layer can be the same as that in Exemplary Embodiment 3.

[0066] Exemplary Example 2

[0067] This exemplary embodiment provides a drilling risk identification model.

[0068] The drilling risk identification model may be obtained by the method for establishing the drilling risk identification model as described in Exemplary Example 1.

[0069] Exemplary Example 3

[0070] This exemplary embodiment provides a method for intelligent identification of drilling risks.

[0071] In an exemplary embodiment of the drilling risk intelligent identification method, the risk intelligent identification method may include the following steps:

[0072] S100: Obtain multiple risk characterization parameters.

[0073] In this embodiment, obtaining multiple risk characterization parameters may include the type of security risk and the sampling time.

[0074] In the case of a risk type of formation gas production risk, the multiple risk characterization parameters include: pipeline pressure relief, methane concentration, pipeline differential pressure, and oxygen concentration.

[0075] In the case of a risk type of formation water outflow risk, the multiple risk characterization parameters include hook load, riser pressure, and outlet humidity.

[0076] In the case of a stuck drill risk, the multiple risk characterization parameters include rotary table speed and rotary table torque.

[0077] S200. Use a convolutional neural network to extract the feature information of each parameter and the relationship between each parameter to obtain composite feature information.

[0078] In this embodiment, the convolutional neural network includes a first convolutional network layer and a second convolutional network layer.

[0079] The first convolutional network layer is used to extract the feature information implicit in each parameter itself.

[0080] The kernel size of the first convolutional network layer is m*1, which represents the time feature information of a certain feature parameter in m rows and 1 column, where m is the vertical length of the convolutional kernel.

[0081] The second convolutional network layer is used to extract the implicit relationships between the parameters.

[0082] The kernel size of the second convolutional network layer is 1*n, which is used to extract the correlation feature information between parameters of a certain risk in a 1 row and n columns, where n is the number of feature parameters of various security risks.

[0083] S300. Classify the composite feature information using an RBF neural network.

[0084] In this embodiment, the RBF neural network is an optimized and trained network.

[0085] The K-means algorithm can be used to initialize the cluster centers and center widths of the RBF neural network.

[0086] Because inappropriate cluster centers may prevent the network from converging effectively during training, the K-means algorithm's cluster centers c can be trained and updated using the loss function component of the mean squared distance error. r The value of can be used to train and update the cluster centers of the K-means algorithm based on the loss function component of the mean squared distance error.

[0087] S400. Real-time risk identification of drilling risks is performed by using the relationship between the number of hidden layer nodes and the evaluation value of different risks, and the probability of risk occurrence of the risk characterization parameter model is obtained.

[0088] In this embodiment, the intelligent recognition method further includes a criterion for selecting the number of hidden layer nodes. This criterion is used to improve the computational efficiency and generalization ability of the model.

[0089] The hidden layer selection rules include: assigning weights to the accuracy and training time of the network model, and determining whether the selection of hidden layer nodes is reasonable based on the ratio of the two parameters corresponding to different numbers of hidden layer nodes.

[0090] In this embodiment, the selection rule for the number of hidden layer nodes may include the following steps:

[0091] (1) Determine the number of input layers of the convolutional network and the number of output layers of the RBF neural network, and determine the basic network model, which includes the convolutional network and the RBF neural network.

[0092] (2) According to Determine the initial number of hidden layer nodes in the RBF neural network, where l is the number of input nodes in the RBF network; r is the number of hidden layer nodes; k is the number of output layer nodes in the RBF network; and ε is a constant coefficient with a value ranging from 0 to 10.

[0093] (3) Train the basic network model to obtain the final training time T and the accuracy of the basic network model test set B. In order to eliminate the influence of the units between the indicators, the training time and the accuracy of the network model test set are normalized to 0-1.

[0094] (4) The evaluation value C(r) corresponding to different numbers of hidden layer nodes is obtained using the following formula, where, μ and v are constant coefficients.

[0095] (5) For different risk types, set different increments of the number of hidden layer nodes Δr, and repeat step (4) to obtain a new evaluation value C(r+1). If the new evaluation value continues to decrease or tends to remain unchanged, stop increasing the number of hidden layer nodes.

[0096] (6) Normalize the evaluation values ​​corresponding to different numbers of hidden layer nodes to 0-1, and take the number of hidden layer nodes corresponding to the point where the normalized evaluation value C(r) is 100% as the optimal number of nodes.

[0097] Exemplary Example 4

[0098] This exemplary embodiment provides a computer-readable storage medium storing a computer program. The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform a method for establishing a drilling risk identification model according to the present invention, or a method for intelligent identification of drilling risks. The computer-readable recording medium is any data storage device capable of storing data readable by a computer system. Examples of computer-readable recording media include: read-only memory, random access memory, read-only optical disk, magnetic tape, floppy disk, optical data storage device, and carrier waves (such as data transmission via the Internet through wired or wireless transmission paths).

[0099] Exemplary Example 5

[0100] This exemplary embodiment provides a computer device. The computer device includes a processor and a memory. The memory is used to store a computer program. The computer program is executed by the processor, causing the processor to perform the computer program of the method for establishing a drilling risk identification model according to the present invention, or the computer program for intelligent identification of drilling risks.

[0101] To better understand the above exemplary embodiments of the present invention, the present invention will be further explained below in conjunction with the accompanying drawings and specific examples, but the present invention is not limited to the following embodiments.

[0102] Example 1

[0103] This example provides an intelligent identification method for drilling risks. For complex drilling risks, this example uses an improved convolutional neural network based on the RBF classifier to achieve fast and accurate identification of drilling risks.

[0104] Compared to traditional fully connected neural networks, RBF neural networks can map low-dimensional data to high-dimensional data, enabling nonlinear classification tasks to a certain extent. Furthermore, RBF neural networks are the best approximation of continuous functions, exhibiting a "local mapping" characteristic, thus resulting in better generalization ability and faster computation speed. Field application results demonstrate that this method has strong adaptability and effectiveness, and can identify the occurrence of abnormal risks earlier than traditional convolutional neural networks.

[0105] like Figure 1 As shown, the risk identification method for gas drilling can be implemented using the following technical solutions.

[0106] Step 1: Analyze and summarize the relevant characterization parameters for three common safety risks during gas drilling monitoring: formation gas production risk, formation water production risk, and stuck pipe risk. Construct a risk characterization parameter model, and use this model as input samples into a convolutional network.

[0107] Because the relevant characterization parameters change rapidly and significantly when formation gas production risk and stuck drill risk occur, a sampling time of 60 seconds can be selected when constructing the parameter model. However, when formation water production risk occurs, the relevant parameters change slowly and with smaller amplitudes. To fully reflect the parameter change trend during formation water production, a sampling time of 120 seconds can be selected.

[0108] Finally, risk characterization parameter models were constructed for the three different risks mentioned above, as shown in Table 1.

[0109] Table 1 Safety Risk Characterization Parameter Model

[0110]

[0111] Step 2: Use the established risk characterization parameter model as the input sample for the convolutional layer. The input sample is a p×n (p=30,60) matrix, where p is the vertical length of the sample and n is the number of various safety risk feature parameters.

[0112] Generally, conventional convolutional neural networks typically use a stacked approach for convolutional layers, with each layer containing only a single kernel of a specific size. This invention, however, employs two convolutional layers—a first convolutional network layer C1 and a second convolutional network layer C2—to acquire feature information at different scales.

[0113] Because the input data model structure is different for each risk, the convolution kernel size used for different risks is also different.

[0114] In this example, the input model of the convolutional layer is m*n dimensional, where m is the time span and n is the number of representation parameters. The convolutional layer built in this example first obtains the feature information of the input data at different time spans through convolutional kernels of different m*1 sizes, and then obtains the feature information between the risk representation parameters through convolutional kernels of 1*n size.

[0115] Specifically, the kernel size of the first convolutional network layer C1 is m*1, which is the time feature information of a certain feature parameter in m rows and 1 column, where m is the vertical length of the convolutional kernel.

[0116] The second convolutional network layer C2 has a convolutional kernel size of 1*n, which is used to extract the correlation feature information between parameters of a certain risk in a 1 row and n column representation parameter, where n is the number of feature parameters of various security risks.

[0117] In addition, because the ELU function can better improve the learning characteristics of convolutional neural networks compared to other activation functions, thereby effectively preventing gradient vanishing and having better generalization performance, the present invention can choose the ELU function as the activation function of the convolutional layer.

[0118] The expression for the ELU function is:

[0119] Step 3: Use a non-linear network as a classifier to classify the data obtained from the convolutional network layers.

[0120] Because the monitoring data exhibit highly complex nonlinear characteristics during the occurrence of abnormal risks in the drilling process, it is necessary to map the stuck pipe risk monitoring data and normal drilling risk monitoring data collected from multiple wells onto a two-dimensional plane PC1-PC2 after feature extraction processing, to obtain, as shown below. Figure 2 The projection results are shown below. It should be noted that in the figure, PC1 and PC2 represent the coordinates of the data points projected onto the two-dimensional plane; circles represent category 1, indicating no risk of stuck drill; triangles represent category 2, indicating a risk of stuck drill.

[0121] Depend on Figure 2 It can be seen that although the rotary table speed and rotary table torque parameters are highly correlated with the risk of stuck drills and can characterize the risk of stuck drills to a certain extent, the two types of risk data points cannot be accurately distinguished directly by a straight line on a two-dimensional plane. That is, there is a certain nonlinear relationship between the rotary table speed and rotary table torque parameters and whether the risk of stuck drills occurs.

[0122] Traditional convolutional neural networks often classify data through fully connected layers. When dealing with such complex nonlinear problems, fully connected layers fit the data by linearly weighting all nodes before each layer.

[0123] Traditional methods for classifying complex nonlinear data typically achieve approximate nonlinear classification by increasing the number of hidden layers or nodes in fully connected layers. However, due to inherent structural limitations, this often fails to yield ideal results. Furthermore, structural complexity leads to an excessive number of parameters, low computational efficiency, and an inability to meet the speed requirements of drilling operations, while also exhibiting low generalization ability.

[0124] Therefore, this invention uses a nonlinear network as a classifier to classify the data.

[0125] This invention can use an RBF neural network to perform data classification. An RBF neural network is a three-layer feedforward neural network consisting of an input layer, hidden layers, and an output layer. Each layer has multiple neurons, and adjacent layers are unidirectionally connected. The information transfer from the input layer to the hidden layer is a non-linear process. The hidden layer typically uses a Gaussian kernel function as its radial basis function. The radial range is controlled by calculating the Euclidean distance between the input vector and the cluster centers and adjusting the width of the center vector. The output of the hidden layer node is then obtained by calculating the Gaussian value of the input vector.

[0126]

[0127] in, A is the output of the r-th hidden layer node, dimensionless; A is the output amplitude of the hidden layer, dimensionless; r is the number of hidden layer nodes, dimensionless.

[0128] Take amplitude value By selecting different center vectors and their widths, a two-dimensional Gaussian kernel function graph is obtained, as shown below. Figure 3 As shown, Figure 3 This demonstrates how different values ​​of cluster centers and center widths are used to observe the effects of cluster centers and widths on the Gaussian function response when the input data range is [-5, 5].

[0129] In the figure, the vertical axis The x-axis represents the Gaussian value corresponding to the input data point x; the x-axis represents the input data interval; the solid line represents the Gaussian value when c = 0. 2 When c = 0.2, the Gaussian function has the strongest response at x = 0, and its range of action is x ∈ (-1.2, 1.2); the dashed line represents the range when c = 0, δ 2 When c = 5, the Gaussian function has the strongest response at x = 0, and its range of action is x ∈ (-5, 5); the point line represents the response when c = -2, δ 2 When x = 0.2, the Gaussian function has the strongest response when x = -2, and its range of action is x ∈ (-3.2, -0.8).

[0130] The Gaussian kernel function, as a nonlinear function, is adjusted... The magnitude of the distance between the weights and the input vector controls the range of the function mapping. When the input signal x is close to the center of the function, the hidden layer nodes will produce a larger output. When the input distance is 0, the function output reaches its maximum value. As the distance between the weights and the input vector increases, the network output decreases non-linearly, meaning that the radial basis function responds locally to the input signal.

[0131] This demonstrates that this network possesses local nonlinear approximation capabilities; that is, by calculating the Gaussian kernel function value of data points, it can map them to a higher-dimensional space. According to Cover's theorem, mapping the input vector from a low-dimensional linearly inseparable case to a higher-dimensional space will make the data linearly separable.

[0132] Choose the center coordinates as (0, 0). pass

[0133] Step 4: Analyze the upgraded 3D spatial data points using an RBF neural network to simulate the internal regularity of the data.

[0134] The upgraded three-dimensional spatial data points are distributed on a Gaussian surface. Compared to two-dimensional space, there is a plane that makes the data closer to the condition of direct linear separability. Although some data points are still not directly separable, their number has been greatly reduced, which also means a significant improvement in classification accuracy.

[0135] If a higher-dimensional space is used, the classification accuracy can be further improved. Regarding the network structure, the closer the input vector x is to the center vector cr, the stronger the response of the hidden layer neurons and the greater their impact on the final output; conversely, the impact is minimal. This fully demonstrates the fast computation speed and local approximation characteristics of the RBF neural network.

[0136] By utilizing the properties of the Gaussian kernel function, the RBF neural network can theoretically approximate any nonlinear function through a simple three-layer network structure, and can simulate the difficult-to-analyze regularities within the data.

[0137] Step 5: Perform the following expression on the output of the hidden layer nodes: The weighted summation yields the output Y = [y1, y2] of the neural network. T , where y k ω represents the output of the output layer node; ω represents the weights of each layer in the neural network; and after activation by the Softmax function, it represents the probability of a certain security risk occurring.

[0138] Based on the above steps, the final neural network model structure diagram of this invention is drawn as follows: Figure 4 As shown.

[0139] Example 2

[0140] This example provides a drilling risk intelligent identification method, which adds an optimization step to the drilling risk intelligent identification method in Example 1: using the K-means algorithm to determine the initial parameters of the RBF neural network model.

[0141] Specifically, the K-means algorithm is used to determine the initial parameters of the RBF neural network model.

[0142] Generally, the cluster centers and widths of an RBF neural network directly affect the final result of the entire model. The K-means algorithm is a typical unsupervised learning method that does not require defined labels for the input data, using only the distance between samples as the similarity criterion, which aligns with the application scenario of this invention. Therefore, this invention utilizes the K-means algorithm to initialize the cluster centers for the feature information identified by the convolutional network.

[0143] The specific steps for optimizing the layer parameters of an RBF neural network model based on the K-means algorithm may include:

[0144] Step 1: Use the composite feature information extracted by the convolutional neural network as the input X = {x1, x2, ..., x} to the K-means algorithm. N}, randomly select h different data as initial cluster centers.

[0145] Step 2: Calculate the input sample X and the initial cluster centers c r Distance between (s): dis(X,c) r (s))=‖Xc r (s)‖r=1,2,…,h, divide the input sample X according to the minimum distance criterion.

[0146] Step 3: Based on dis(X,c) r (s))=‖Xc r The input sample X corresponding to the minimum Euclidean distance calculated by (s)‖r=1,2,…,h is assigned to cluster λ. r In (s), that is

[0147] Step 4: Repeat steps 2 and 3 until the new cluster centers no longer change, then output the final cluster centers c. r According to the formula, Calculate the width δ of the hidden layer nodes r ,in Let c be the distance between the r-th cluster center and the other nearest sample centers, and finally obtain the initial cluster center c. r and initial center width δ r .

[0148] Example 3

[0149] This example provides a drilling risk intelligent identification method, which adds a step to the drilling risk intelligent identification method in Example 2: training the entire network model based on the loss function component of the distance mean square error, thereby better optimizing the model parameters.

[0150] Specifically, the K-means algorithm determines the initial center point and width of the network based on the characteristics of the training samples. However, during the training process of the network model of this invention, the feature information extracted by the convolutional neural network layer will be continuously optimized, and the optimal cluster center may change. This will cause the cluster center and width initially initialized by the K-means algorithm to become unsuitable.

[0151] To achieve faster training efficiency and more accurate recognition results, this invention proposes a loss function component based on the mean squared error of distance to train the entire network model, thereby better optimizing the model parameters.

[0152] Specifically, the following steps are included:

[0153] Step 1: Based on the input sample X and the initialized cluster centers c r Distance between (s): dis(X,c) r (s))=‖Xc r (s)‖r=1,2,…,h, yielding the loss function components based on the mean squared error of the distance:

[0154]

[0155] Where, loss un sup ervised This involves summing the mean squared errors of the distances between each sample and the cluster center.

[0156] Step 2: Modify the calculation method of the loss function during network model training according to the above formula as follows:

[0157] Loss = loss mean-squared +loss unsuoervised ,

[0158] Where, loss mean-squared Let be the mean squared error loss function.

[0159] Step 3: The calculated results are backpropagated and trained using the Adam optimizer. Simultaneously, all weights and biases of the entire neural network are optimized, and c is also optimized. r and δ r , where c r Let δ be the center vector of the r-th hidden layer node; r Let be the width of the center vector of the r-th hidden layer node.

[0160] Example 4

[0161] This example provides a drilling risk intelligent identification method that adds a step to the drilling risk intelligent identification method in Examples 1, 2 or 3: a new hidden layer node number selection step.

[0162] Specifically, under the condition of a limited number of hidden layers, the number of hidden layer nodes in an RBF neural network has a significant impact on the overall network performance. Too few hidden layer nodes can lead to underfitting, while too many hidden layer nodes can make the model structure too complex and reduce generalization ability. In drilling risk intelligent identification scenarios, both computational efficiency and generalization ability of the model are required to be high. Therefore, in order to meet the needs of field applications, the appropriate selection of the number of hidden layer nodes becomes extremely important.

[0163] This example proposes a method for selecting the number of hidden layer nodes. This method assigns weights to the network model's test set accuracy and training time, and determines whether the selection of hidden layer nodes is reasonable based on the ratio of the two parameters corresponding to different numbers of hidden layer nodes.

[0164] Specifically, this may include the following steps:

[0165] Step 1: Determine the number of input and output layers in the network model to establish the basic model structure.

[0166] Step 2: According to Select an initial number of hidden layer nodes, where l is the number of input nodes of the RBF network, r is the number of hidden layer nodes, k is the number of output layer nodes of the RBF network, and ε is a constant coefficient from 0 to 10.

[0167] Step 3: Train the network model to obtain the final training time T and the network model test set accuracy B. In order to eliminate the influence of the units between the indicators, the training time and the network model test set accuracy are normalized to 0-1.

[0168] Step 4: Assign weights μ and v based on the importance of the training time and the accuracy of the network model on the test set, respectively, to obtain the evaluation value C(r) corresponding to different numbers of hidden layer nodes. μ and ν are constant coefficients; T is the final training time; B is the network model's accuracy on the test set.

[0169] Step 5: Set different increments Δr for the number of hidden layer nodes for different risk identification models, and repeat step 4 to obtain a new evaluation value C(r+1). If the new evaluation value continues to decrease or tends to remain unchanged, stop increasing the number of hidden layer nodes.

[0170] Specifically, the evaluation value is a ratio between accuracy and training time. The desired result is high accuracy with short training time. Therefore, once the evaluation value reaches 100%, the accuracy will not improve significantly with the increase of the number of hidden layer nodes, but the model will become more complex, leading to a gradual increase in training time. Thus, there is no situation where the evaluation value continuously increases; it will inevitably reach its peak and then gradually decrease.

[0171] Step 6: Normalize the evaluation values ​​corresponding to different numbers of hidden layer nodes to 0-1, and take the number of hidden layer nodes corresponding to the point where the normalized evaluation value C(r) is 100% as the optimal number of nodes.

[0172] For example: Under the hidden layer selection rules shown in this example, the relationship between the evaluation values ​​obtained from the three risks and the number of hidden layer nodes is as follows: Figure 5A , Figure 5B and Figure 5C As shown.

[0173] from Figure 5A It can be seen that the formation gas production model achieves an evaluation value of 100% when the number of hidden layer nodes is 29. The incremental value of the number of hidden layer nodes can be selected as 1. After the evaluation result reaches the optimal value, the evaluation result gradually deteriorates as the number of hidden layer nodes increases.

[0174] from Figure 5B It can be seen that the formation water discharge model achieves an evaluation value of 100% when the number of hidden layer nodes is 40. The optimal evaluation result can be achieved by selecting an increment of 5 for the number of hidden layer nodes. After reaching the optimal evaluation result, the evaluation result gradually deteriorates as the number of hidden layer nodes increases.

[0175] from Figure 5C It can be seen that the evaluation value of the drill bit model reaches 100% when the number of hidden layer nodes is 25. The increment of the number of hidden layer nodes can be selected as 1. After the evaluation result reaches the optimal level, the evaluation result gradually deteriorates as the number of hidden layer nodes increases.

[0176] Example 5

[0177] This example provides a drilling risk identification model, which is built upon the drilling risk intelligent identification methods in Examples 1, 2, 3, or 4.

[0178] This example shows the results of training using monitoring data collected from gas drilling monitoring systems in different regions.

[0179] In this example, the drilling risk identification model of the present invention is trained 8000 times simultaneously using a convolutional neural network and an RBF neural network.

[0180] The data in this example uses monitoring data collected by gas drilling monitoring systems in different regions. The historical monitoring data of more than ten wells are organized according to the relevant characterization parameters when formation gas production, formation water production, and stuck pipe risk occur. Any one of the drilling risk intelligent identification methods in Examples 1, 2, 3, or 4 can be used to construct the sample model and build the classification network.

[0181] The drilling risk identification model proposed in this example is compared with the traditional convolutional neural network. In order to verify whether the drilling risk identification model in this example is better than a single RBF neural network, an RBF network model is constructed in the same way for performance comparison.

[0182] Figure 6A and Figure 6B The training results of the sample training set for the formation gas production risk in this example are shown.

[0183] Figure 6AThe diagram shows the loss values ​​of each neural network after training. In the figure, solid lines represent the loss values ​​of the drilling risk identification model of this invention, dashed lines represent the loss values ​​of the convolutional neural network, and dotted lines represent the loss values ​​of the RBF neural network. Figure 6A It can be seen that the drilling risk identification model of the present invention has the best training effect, and the loss value converges after 3000 training iterations.

[0184] Figure 6B The figure shows the accuracy of each neural network on the training set after training. In the figure, the solid line represents the accuracy of the drilling risk identification model of the present invention, the dashed line represents the accuracy of the convolutional neural network, and the dotted line represents the accuracy of the RBF neural network. Figure 6B It can be seen that the accuracy of the final training set of the drilling risk identification model of the present invention reaches 98%.

[0185] Figure 7A and Figure 7B The training results of the sample training set for formation water outflow risk in this embodiment are shown.

[0186] Figure 7A The diagram shows the loss values ​​of each neural network after training. In the figure, solid lines represent the loss values ​​of the drilling risk identification model of this invention, dashed lines represent the loss values ​​of the convolutional neural network, and dotted lines represent the loss values ​​of the RBF neural network. Figure 7A It can be seen that the loss values ​​of the three neural networks eventually converged after 8000 training iterations.

[0187] Figure 7B The figure shows the accuracy of each neural network on the training set after training. In the figure, the solid line represents the accuracy of the drilling risk identification model of the present invention, the dashed line represents the accuracy of the convolutional neural network, and the dotted line represents the accuracy of the RBF neural network. Figure 7B It can be seen that the RBF neural network has an accuracy of only 83% on the training set of formation water production, while the convolutional neural network has an accuracy of 94% on the training set. However, the improved drilling risk identification model of this invention has an accuracy of 95% on the training set after 4000 training steps, which remains basically unchanged.

[0188] The parameters characterizing the risk of stuck drill pipe include two parameters: rotary table speed and rotary table torque. The training results of the sample training sets for the three networks on the risk of stuck drill pipe are as follows: Figure 8A and Figure 8B As shown.

[0189] Figure 8A The diagram shows the loss values ​​of each neural network after training. In the figure, solid lines represent the loss values ​​of the drilling risk identification model of this invention, dashed lines represent the loss values ​​of the convolutional neural network, and dotted lines represent the loss values ​​of the RBF neural network. Figure 8A It can be seen that the loss values ​​of the three neural networks converge very quickly.

[0190] Figure 8B The figure shows the accuracy of each neural network on the training set after training. In the figure, the solid line represents the accuracy of the drilling risk identification model of the present invention, the dashed line represents the accuracy of the convolutional neural network, and the dotted line represents the accuracy of the RBF neural network. Figure 8B It can be seen that all three neural networks have good recognition performance on the training set samples, with an accuracy rate of over 90%.

[0191] The accuracy of the test set on the three neural networks is shown in Table 3. The network training loss values ​​converged under different risks, indicating that all three neural networks have good training effects. However, the final recognition results on the test set show differences in the recognition accuracy of the three neural networks. Table 3 shows that the improved drilling risk identification model of this invention achieves accuracy rates of 97.32%, 95.25%, and 93.78% on the test sets for the three risks, respectively, all of which are better than the test results of the convolutional neural network and the RBF neural network.

[0192] Table 3 Comparison of Test Results for Test Set

[0193]

[0194] The results of this example demonstrate that, in the context of drilling risk identification, the proposed drilling risk identification model exhibits superior performance, successfully inferring the inherent regularities within the data. The use of multiple convolutional kernels to simultaneously extract feature information in the convolutional layers enhances the network model's ability to extract richer features from the data. Applying the RBF neural network as a classifier to the convolutional neural network effectively strengthens the network model's ability to analyze and process the feature information inherent in the data.

[0195] Example 6

[0196] In this example, the drilling risk intelligent identification method of Examples 1, 2, 3 or 4, or the drilling risk identification model of Example 5, can be used to process the field drilling data.

[0197] This example illustrates the field test results obtained by applying any of the above-described intelligent drilling risk identification methods or drilling risk identification models of the present invention during the formation gas production period of Fenggu A well and Xin A well during the drilling monitoring process.

[0198] Generally, drilling site conditions are complex. Operations such as shutting off dust suppression water or changing the gas injection rate can affect changes in formation gas production risk indicators, especially when methane concentration changes are small, making it difficult for on-site monitoring personnel to react immediately. To reduce the probability of false alarms, the risk identification threshold is set at 60%.

[0199] The field test results of the formation gas production period during the drilling monitoring process of Fenggu A well and Xin A well are selected for illustration. The data of Fenggu A well is shown in Table 4.

[0200] Table 4. Field Application Effects of Formation Gas Production Risk Assessment in Well Fenggu A

[0201]

[0202] As can be seen from Table 4, during the gas production period of Fenggu A well, the drilling risk intelligent identification method of the present invention determined the formation gas production at 22:06:45, while the CNN convolutional neural network and RBF network models only identified the formation gas production at 22:07:25 and 22:07:29, respectively.

[0203] The network model determined formation gas production in well A at 11:38:59, while the CNN convolutional neural network and RBF network models only identified formation gas production at 11:39:17 and 11:39:15, respectively. Field application results of the formation gas production risk identification method demonstrate that the intelligent drilling risk identification method of this invention has a rapid response and significantly outperforms the CNN convolutional neural network and RBF neural network, and shows good results in gas drilling scenarios in different regions.

[0204] The test results of formation water production during the monitoring while drilling of wells Xin A and Xin Sheng B are used for illustration. The testing method and the method for determining whether the risk has occurred are the same as the method for identifying formation gas production risks. The test results of well Xin A are shown in Figure 5.

[0205] Table 5. Field Application Effects of Formation Water Outflow Risk Assessment in Well A

[0206]

[0207] As shown in Table 5, during the formation water emergence stage of Well A, the drilling risk intelligent identification method of the present invention determined the formation water emergence risk at 10:21:19, even before the on-site outlet humidity changed significantly, based on the changes in multiple characterization parameters. This was 8 seconds earlier than the CNN convolutional neural network and 22 seconds earlier than the RBF network.

[0208] In a field application at another well site, Xinsheng B well, the intelligent drilling risk identification method of this invention determined formation water inflow at 5 hours, 41 minutes, and 51 seconds, 16 seconds earlier than the identification results of CNN convolutional neural network and RBF network. This demonstrates that the network model proposed in this invention can effectively and promptly assess formation water inflow risk based on the characteristic parameters at the time of risk occurrence, achieving a better identification effect.

[0209] The risk of stuck drill pipe often directly affects whether gas drilling can continue to proceed normally. Therefore, timely identification of stuck drill pipe risk is very important. The intelligent drilling risk identification method of the present invention reacts quickly and can provide timely auxiliary judgment to the monitoring personnel, notify the driller to handle the situation, and ensure the smooth progress of the entire drilling process.

[0210] For example, during the stuck drill risk at Fenggu A well, the intelligent drilling risk identification method of this invention alerted the system at 21:57:37 on-site time, which was 8 seconds and 7 seconds earlier than the identification results of CNN convolutional neural network and RBF neural network, respectively.

[0211] Example 7

[0212] like Figure 9 As shown, a computer device 100 includes a memory 110 and a processor 120. The memory 110 stores a computer program. When the computer program is executed by the processor, it implements the drilling risk identification model establishment method as described above, or the drilling risk intelligent identification method as described above.

[0213] Although the present invention has been described above in conjunction with exemplary embodiments and accompanying drawings, those skilled in the art should understand that various modifications can be made to the above embodiments without departing from the spirit and scope of the claims.

Claims

1. A method for establishing a drilling risk identification model, characterized in that, The method includes the following steps: Establish a database containing risk characterization parameters; Based on the database, a composite network model is established, which includes a convolutional neural network layer and an RBF neural network layer. The convolutional neural network layer includes a first convolutional layer and a second convolutional layer. The first convolutional layer can extract the feature information implicit in each parameter itself, and the second convolutional layer can extract the relationship implicit in each parameter and combine feature information of different scales to obtain composite feature information. The RBF neural network layer can classify the composite feature information. The composite network model is trained to obtain the drilling risk identification model; In the training process, a criterion for selecting the number of hidden layer nodes was established to improve the computational efficiency and generalization of the model. The selection rules for the number of hidden layer nodes include: assigning weights to the accuracy of network model training and the training time of network model, and determining whether the selection of hidden layer nodes is reasonable based on the ratio of the two parameters corresponding to different numbers of hidden layer nodes. The rules for selecting the number of hidden layer nodes include: (1) Determine the number of input layers of the convolutional network and the number of output layers of the RBF neural network, and determine the basic network model, which includes the convolutional network and the RBF neural network; (2) According to Determine the initial number of hidden layer nodes in the RBF neural network, where, l Input the number of nodes for the RBF network; r This represents the number of hidden layer nodes. k This represents the number of output layer nodes in the RBF network. This is a constant coefficient, with a value range of 0-10; (3) Train the basic network model to obtain the final training time. and the accuracy of the basic network model test set To eliminate the influence of different dimensions between indicators, the training time and the network model test set accuracy are normalized to 0-1. (4) Use the following formula to obtain the evaluation value corresponding to different numbers of hidden layer nodes. ,in, , and These are constant coefficients; (5) Set different increments for the number of hidden layer nodes for different risk types. Repeat step (4) to obtain a new evaluation value. If the new evaluation value continues to decrease or tends to remain unchanged, then stop increasing the number of hidden layer nodes; (6) Normalize the evaluation values ​​corresponding to different numbers of hidden layer nodes to 0-1, and then normalize the evaluation values. The optimal number of nodes is the number of hidden layer nodes corresponding to 100% of the points.

2. The method for establishing a drilling risk identification model according to claim 1, characterized in that, Both the first and second convolutional layers used three convolutional kernels of different sizes.

3. The method for establishing a drilling risk identification model according to claim 1, characterized in that, After establishing the composite network model, the method also includes a step of optimizing the model.

4. The method for establishing a drilling risk identification model according to claim 3, characterized in that, The optimization includes optimization methods for the K-means algorithm.

5. The method for establishing a drilling risk identification model according to claim 4, characterized in that, The training includes: training and updating the cluster centers of the K-means algorithm based on the loss function components of the distance mean square error.

6. A drilling risk identification model, said model being obtained by the method of any one of claims 1 to 5.

7. A method for intelligent identification of drilling risks, characterized in that, The method includes the following steps: Obtain multiple risk characterization parameters; The feature information of each parameter and the relationship between each parameter are extracted using a convolutional neural network to obtain composite feature information. The composite feature information is classified using an RBF neural network; By using the relationship between the number of hidden layer nodes and the evaluation value for different risks, drilling risks are identified in real time, and the probability of risk occurrence of the risk characterization parameter model is obtained. In the intelligent recognition method, a criterion for selecting the number of hidden layer nodes is established to improve the computational efficiency and generalization of the model; The selection rules for the number of hidden layer nodes include: assigning weights to the accuracy of network model training and the training time of network model, and determining whether the selection of hidden layer nodes is reasonable based on the ratio of the two parameters corresponding to different numbers of hidden layer nodes. The rules for selecting the number of hidden layer nodes include: (1) Determine the number of input layers of the convolutional network and the number of output layers of the RBF neural network, and determine the basic network model, which includes the convolutional network and the RBF neural network; (2) According to Determine the initial number of hidden layer nodes in the RBF neural network, where, l Input the number of nodes for the RBF network; r This represents the number of hidden layer nodes. k This represents the number of output layer nodes in the RBF network. This is a constant coefficient, with a value range of 0-10; (3) Train the basic network model to obtain the final training time. and the accuracy of the basic network model test set To eliminate the influence of different dimensions between indicators, the training time and the network model test set accuracy are normalized to 0-1. (4) Use the following formula to obtain the evaluation value corresponding to different numbers of hidden layer nodes. ,in, , and These are constant coefficients; (5) Set different increments for the number of hidden layer nodes for different risk types. Repeat step (4) to obtain a new evaluation value. If the new evaluation value continues to decrease or tends to remain unchanged, then stop increasing the number of hidden layer nodes; (6) Normalize the evaluation values ​​corresponding to different numbers of hidden layer nodes to 0-1, and then normalize the evaluation values. The optimal number of nodes is the number of hidden layer nodes corresponding to 100% of the points.

8. The intelligent identification method for drilling risks according to claim 7, characterized in that, The RBF neural network is an optimized and trained network; wherein, The optimization includes: initializing the cluster centers and center widths of the RBF neural network using the K-means algorithm; The training includes: training and updating the cluster centers of the K-means algorithm based on the loss function components of the distance mean square error.

9. The intelligent identification method for drilling risks according to claim 7, characterized in that, The acquisition of multiple risk characterization parameters includes security risk type and sampling time, wherein, In the case of risk type being formation gas production, the multiple risk characterization parameters include: pipeline pressure relief, methane concentration, pipeline differential pressure, and oxygen concentration. In the case of a risk type of formation water outflow, the multiple risk characterization parameters include hook load, riser pressure, and outlet humidity; In the case of a stuck drill risk, the multiple risk characterization parameters include rotary table speed and rotary table torque.

10. The risk intelligent identification method according to claim 7, characterized in that, The convolutional neural network comprises a first convolutional network layer and a second convolutional network layer, wherein... The first convolutional network layer is used to extract the feature information implicit in each parameter of the parameter model itself; The second convolutional network layer is used to extract the implicit relationships between the parameters in the parametric model.

11. The risk intelligent identification method according to claim 10, characterized in that, The kernel size of the first convolutional network layer is m*1, which is the time feature information of a certain feature parameter in m rows and 1 column, where m is the vertical length of the convolutional kernel; The kernel size of the second convolutional network layer is 1*n, which is used to extract the correlation feature information between parameters of a certain risk in a 1 row and n columns, where n is the number of feature parameters of various security risks.

12. A computer device, characterized in that, The device includes: processor; The memory stores a computer program, which, when executed by a processor, implements the method for establishing a drilling risk identification model as described in any one of claims 1 to 5, or implements the intelligent identification method for drilling risks as described in any one of claims 7 to 11.

13. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for establishing a drilling risk identification model as described in any one of claims 1 to 5, or the method for intelligent identification of drilling risks as described in any one of claims 7 to 11.

Images