Small-Sample Condition Monitoring Method and Device for Drilling Process Based on Data Augmentation

Abstract

The present invention provides a small-sample status monitoring method and device for the drilling process based on data augmentation. The drilling process data is preprocessed to ensure the quality and efficiency of data generation; an improved generative adversarial network is used to train a data generation model by combining a small amount of abnormal drilling data from the target well and the reference well. According to the characteristics of the drilling process, the generative adversarial network is improved mainly in two aspects: the design of the model parameter configuration module and the construction of the feature distance loss function; based on the trained data generation model and the reference well data, the abnormal drilling data of the target well is augmented to achieve data augmentation; an equal proportion of actual normal drilling data and synthetic abnormal drilling data are used to establish a monitoring model to achieve the status monitoring of the drilling process under small-sample conditions. By improving the generative adversarial network, the sample data set in the abnormal drilling state is augmented, the generalization ability of the drilling status monitoring model is improved, the monitoring effect is improved, and safety guarantee is provided for the drilling process.

Description

Technical Field

[0001] The present invention relates to the field of geological drilling engineering, and particularly to a small-sample status monitoring method and device for the drilling process based on data augmentation. Background Art

[0002] In recent years, China's demand for resources and energy has been increasing continuously. In 2019, the total primary energy production was 3.97 billion tons of standard coal, and the total consumption was 4.86 billion tons of standard coal, increasing by 5.1% and 3.3% respectively compared with the previous year. It is estimated that by 2030, the external dependence of minerals such as oil, iron ore, and copper will reach about 70%, 85%, and 80% respectively. At the same time, China has great potential for deep resource and energy development, and it is practical to obtain resources and energy from the deep. Defining deep drilling according to the exploration depth of resources, about 40% of the remaining oil resources in China are deep-layer oil resources, and about 60% of the remaining natural gas resources are deep-layer natural gas resources. Therefore, deep geological exploration and development have become inevitable.

[0003] During the deep geological exploration process, due to factors such as complex and changeable geomechanical environments and fatigue and aging of drilling tools, various drilling accidents exist. Therefore, it is necessary to carry out status monitoring during the drilling process.

[0004] The data-driven status monitoring method does not require the establishment of a complex mechanism model, has a wide adaptation range and high portability, and has been widely used in industrial processes. However, this type of method is greatly affected by data, and the small-sample problem will lead to insufficient model training and poor generalization ability. During the drilling process, this problem is more prominent. The existing methods mainly use the accident data of other boreholes to establish a drilling status monitoring model. However, due to the different geological conditions of the regions where each wellbore is located and the different equipment used for drilling, there are differences in the data distribution among the boreholes, and this distribution difference will have a negative impact on the application of the model. Therefore, it is necessary to propose a drilling process status monitoring method suitable for small-sample conditions. Summary of the Invention

[0005] During the drilling process, abnormal status data is relatively scarce, and when applying the data-driven drilling status monitoring method, there are problems of insufficient training and poor model generalization ability. The existing methods are easily affected by the data distribution differences among different drillings and are difficult to achieve good application effects. To solve the above problems, according to the multi-time scale and multi-variable characteristics of the drilling process, the present invention improves the generative adversarial network, and thus provides a small-sample status monitoring method and device for the drilling process based on data augmentation.

[0006] According to one aspect of the present invention, a small-sample status monitoring method for the drilling process based on data augmentation mainly includes the following steps:

[0007] S1: For the drilling data of the reference well and the drilling data of the target well Perform feature selection preprocessing to screen out features that are highly relevant to the drilling process;

[0008] S2: Considering the multi-time scale and multi-variable characteristics of the drilling process, improve the generative adversarial network by designing a model parameter configuration module and constructing a feature distance loss function; Based on the improved generative adversarial network, use the preprocessed abnormal drilling data of the reference well in step S1 and the abnormal drilling data of the target well to train and obtain a data generation model ;

[0009] S3: Use the abnormal drilling data of the reference well as the input of the data generation model , and the output of the data generation model is the synthetic abnormal drilling data of the target well ;

[0010] S4: Select a base classifier, and combine the balanced drilling data set, including an equal proportion of the synthetic abnormal drilling data of the target well and the actual normal drilling data , to establish a drilling state monitoring model to realize the drilling process state monitoring under small sample conditions.

[0011] Furthermore, step S1 includes:

[0012] Adopt the chi-square test method to perform feature selection preprocessing on the drilling data of the reference well and the drilling data of the target well , and screen out features whose chi-square statistic is greater than the probability threshold.

[0013] Furthermore, the step of adopting the chi-square test method to perform feature selection preprocessing on the drilling data of the reference well and the drilling data of the target well , and screening out features whose chi-square statistic is greater than the probability threshold includes:

[0014] S1.1: Calculate the chi-square statistic of each feature in the drilling data of the reference well and the drilling data of the target well , and the calculation formula is:

[0015]

[0016] In the formula, is the observed frequency of the actual data, is the expected frequency, K is the product of the number of test hypotheses and the number of drilling state categories,k = 1, 2, ..., K ;

[0017] S1.2: Compare the chi-square statistics of each feature with the probability threshold T . If is greater than T , retain the corresponding feature; otherwise, remove the corresponding feature.

[0018] Furthermore, step S1 includes:

[0019] Adopt the Xgboost method to perform feature selection preprocessing on the drilling data of the reference well and the drilling data of the target well to screen out the features highly correlated with the drilling process.

[0020] Furthermore, step S2 includes:

[0021] S2.1: Construct a generative adversarial network, where the generator G is used to generate synthetic drilling data, and the discriminator D is used to distinguish between synthetic data and real data. By minimizing the generator loss and maximizing the discriminator loss, complete the network training, G and D when reaching equilibrium, G is the data generation model obtained through training ;

[0022] Given m abnormal drilling samples of the target well , and m abnormal drilling samples of the reference well , use the sample as the input of the generator G , and use the sample and the output of the generator as the input of the discriminator D . Then the objective function is expressed as:

[0023]

[0024] In the formula, is the data distribution of the target well, is the data distribution of the reference well, is the output of the generator, is the output of the discriminator, is the discriminator loss, is the generator loss;

[0025] S2.2: For mutation types and slow-varying types For drilling accidents, different initial values are set for the training parameters of the data generation model, including the learning rate lr and the batch size bs :

[0026]

[0027] In the formula, is the drilling accident category label, and are the parameter combinations corresponding to the mutation type and the slow change type respectively;

[0028] S2.3: Update the network parameters by means of stochastic gradient descent; Calculate the magnitude and direction of the gradient descent according to the loss function, and the generator loss function and the discriminator loss function are respectively:

[0029]

[0030]

[0031] In the formula, is the synthetic data sequence corresponding to the f th feature, is the real data sequence corresponding to the f th feature, r is the number of features after feature selection, is the loss between the two sequences (used as the distance between features), m is the number of samples participating in the network training.

[0032] Furthermore: In step S4, the base classifier includes at least one of: a probabilistic neural network, an artificial neural network, and a convolutional neural network.

[0033] According to another aspect of the present invention, a small-sample state monitoring device for the drilling process based on data augmentation includes the following modules:

[0034] A preprocessing module for performing feature selection preprocessing on the reference well drilling data and the target well drilling data to screen out the features with high correlation with the drilling process;

[0035] A generation model training module for improving the generative adversarial network by designing a model parameter configuration module and constructing a feature distance loss function according to the multi-time scale and multi-variable characteristics of the drilling process; Based on the improved generative adversarial network, using the preprocessed abnormal drilling data of the reference well and the abnormal drilling data of the target well Train a data generation model ;

[0036] An abnormal drilling data generation module for using the abnormal drilling data of the reference well as the input of the data generation model, and the output of the data generation model is the synthetic abnormal drilling data of the target well ; ;

[0037] A drilling state monitoring module for selecting a base classifier and combining the balanced drilling data set, including equal proportions of the synthetic abnormal drilling data of the target well and the actual normal drilling data , to establish a drilling state monitoring model and realize the drilling process state monitoring under small sample conditions.

[0038] The beneficial effects brought by the technical solution provided by the present invention are: preprocessing the drilling process data to ensure the quality and efficiency of data generation; using an improved generative adversarial network and combining a small amount of abnormal drilling data of the target well and the reference well to train the data generation model. According to the characteristics of the drilling process, the generative adversarial network is improved mainly in two aspects: the design of the model parameter configuration module and the construction of the feature distance loss function; based on the trained data generation model and the reference well data, the abnormal drilling data of the target well is expanded to achieve data enhancement; using equal proportions of the actual normal drilling data and the synthetic abnormal drilling data to establish a monitoring model to realize the drilling process state monitoring under small sample conditions. By improving the generative adversarial network, expanding the sample data set in the abnormal drilling state, improving the generalization ability of the drilling state monitoring model, improving the monitoring effect, providing safety guarantee for the drilling process, and having practicability and applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0040] Figure 1 is the block diagram of the small sample state monitoring method for the drilling process based on data enhancement in the embodiment of the present invention;

[0041] Figure 2 is the schematic diagram of the training principle of the drilling data generation model in the embodiment of the present invention;

[0042] Figure 3 is the schematic diagram of abnormal drilling data generation in the embodiment of the present invention;

[0043] Figure 4 is the bar chart of the chi-square statistic corresponding to each feature in the embodiment of the present invention;

[0044] Figure 5Data generation effects of two types of accidents under different initial parameter combinations in the embodiments of the present invention;

[0045] Figure 6 Variation trends of the losses of the generator and discriminator in the embodiments of the present invention;

[0046] Figure 7 Comparison between real data and synthetic data in the case of stuck pipe accidents in the embodiments of the present invention, (a) is real data, and (b) is synthetic data;

[0047] Figure 8 Comparison between the torque sequences (a) of real data and synthetic data and their cumulative distribution functions (b) in the embodiments of the present invention;

[0048] Figure 9 Structural diagram of the small sample condition monitoring device for the drilling process based on data augmentation in the embodiments of the present invention. Detailed implementation manners

[0049] For a clearer understanding of the technical features, objectives, and effects of the present invention, the specific implementation manners of the present invention will now be described in detail with reference to the accompanying drawings.

[0050] The embodiments of the present invention provide a small sample condition monitoring method for the drilling process based on data augmentation.

[0051] Please refer to Figure 1 , Figure 1 which is a flowchart of a small sample condition monitoring method for the drilling process based on data augmentation in the embodiments of the present invention, and specifically includes the following steps:

[0052] S1: Pretreatment of drilling process data. Using the chi-square test method, the drilling data of the reference well ( A Well) and the drilling data of the target well ( Well) B are subjected to feature selection processing to screen out the features with greater correlation with the drilling state, ensuring the quality and efficiency of data generation. The selected features include the normal drilling data and abnormal drilling data of Well A and Well B, and the abnormal drilling data is subsequently used to train a data generation model; The specific process of using the chi-square test for feature selection of drilling process data is as follows:

[0053] S1.1:

[0054] S1.1: is the f th feature in the sample dataset, V is the total number of features. By calculating the chi-square statistic of each feature, the features with greater correlation with the drilling state are screened out. The calculation formula of the chi-square statistic is

[0055]

[0056] Wherein, is the observed frequency of the actual data, is the expected frequency, K is the product of the number of test hypotheses and the number of drilling state categories.

[0057] S1.2: By comparing the chi-square statistic of each feature with the probability threshold T , feature selection is achieved. Judge whether is greater than T . If so, retain the corresponding feature; otherwise, remove the corresponding feature. Determine the value of T by querying the chi-square critical value table. In this embodiment, it is preferable to use the probability value with a significance level of 0.1 and a degree of freedom of 1 as the threshold T .

[0058] As an alternative implementation, the Xgboost method can also be selected to perform feature selection preprocessing on the drilling data of the target well and the reference well and to screen out the features that are highly correlated with the drilling process.

[0059] S2: Train a data generation model based on an improved generative adversarial network. Considering the multi-time scale and multi-variable characteristics of the drilling process, the generative adversarial network is improved by designing a model parameter configuration module and constructing a feature distance loss function. On this basis, use the preprocessed A abnormal drilling data of the well and B abnormal drilling data of the well to train the data generation model ;

[0060] The process of training a data generation model based on an improved generative adversarial network is as follows:

[0061] S2.1: Construction of a generative adversarial network. The generator G is used to generate synthetic drilling data, and the discriminator D is used to distinguish between synthetic data and real data. By minimizing the generator loss and maximizing the discriminator loss, network training is completed, G and D when reaching equilibrium G is the trained data generation model .

[0062] Given m number of B abnormal drilling samples of the well , andm one A Abnormal drilling samples of the well Take the samples as the input of the generator G The samples and the generator output are jointly used as the input of the discriminator D , then the objective function is expressed as:

[0063]

[0064] In the formula is the data distribution of the target well ( B well), is the data distribution of the reference well ( A well), is the generator output, is the discriminator output, is the discriminator loss, is the generator loss;

[0065] S2.2: Model parameter configuration. According to the duration of accident evolution, drilling accidents can be divided into two categories: mutation type and slow - change type, which are respectively expressed as and . In the mutation type, the data mutates in a short time, and the window length of data change is short; in the slow - change type, the data changes slowly over a long time, and the window length is long. To improve the model training accuracy, different initial values of model training parameters are considered for different types of drilling accidents, including the learning rate lr and the batch size bs .

[0066]

[0067] In the formula is the drilling accident category label, and are the parameter combinations corresponding to the mutation type and the slow - change type respectively.

[0068] S2.3: Loss function construction. After giving the initial values of network training parameters, the network parameters are updated by the method of stochastic gradient descent. Calculate the magnitude and direction of gradient descent according to the loss function. The generator loss function and the discriminator loss function are respectively

[0069]

[0070]

[0071] In the formula is thef The synthetic data sequence corresponding to a feature is the true data sequence corresponding to the f th feature r is the number of features after feature selection is the loss between two sequences, also known as the feature distance loss m is the number of samples participating in network training

[0072] S2.4: Model training. As Figure 2 shown, use the preprocessed A abnormal drilling data of well and B abnormal drilling data of well to train and obtain a data generation model .

[0073] S3: B Expansion of the abnormal drilling state sample data set of the well. Based on the trained data generation model , take the A abnormal drilling data of well as the input of the generation model , and the output of the model is B synthetic abnormal drilling data of well ;

[0074] As Figure 3 shown, the process of using the data generation model to expand the abnormal drilling state sample data set of the well is as follows: B Given

[0075] a n number of A abnormal drilling data samples of the well , based on the trained data generation model , take the as the input into the model , and obtain B synthetic abnormal drilling data of well . On this basis, construct B a data set with equal proportions of synthetic abnormal drilling data and normal drilling data of the well as the training set of the drilling state monitoring model

[0076] S4: Drilling process state monitoring based on synthetic data. Use a probabilistic neural network (PNN) as the base classifier, combined with the balanced drilling data set, including equal proportions of B synthetic abnormal drilling data of well and actual normal drilling data , a drilling state monitoring model is established to achieve the drilling process state monitoring under small sample conditions.

[0077] The specific process of the drilling process state monitoring based on synthetic data is as follows:

[0078] For the problem of data-driven drilling state monitoring under small sample conditions, B well synthetic abnormal drilling data and actual normal drilling data are used to establish a drilling process state monitoring model in combination with PNN. PNN can better handle the non-linear mapping problem under multi-feature input and is suitable for the drilling process with multi-variable characteristics. Based on the above description, the drilling process state monitoring under small sample conditions can effectively improve the monitoring performance compared with the traditional modeling method.

[0079] As an optional implementation method, an artificial neural network, a convolutional neural network, etc. can also be selected as the base classifier.

[0080] In this implementation, the drilling data of two wells are selected as the specific object, which involves three drilling states: normal, lost circulation, and stuck pipe. There are a total of 180 groups of data in the dataset, and the ratio of normal drilling data to abnormal drilling data is 6:1. On this basis, the chi-square test method is used to perform feature selection preprocessing on the drilling process data to ensure the quality and efficiency of data generation; an improved generative adversarial network is used to train the data generation model in combination with a small amount of abnormal drilling data from the target well and the reference well. According to the characteristics of the drilling process, the generative adversarial network is improved mainly in two aspects: the design of the model parameter configuration module and the construction of the feature distance loss function; based on the trained data generation model and the reference well data, the abnormal drilling data of the target well is expanded to achieve data enhancement; a monitoring model is established using an equal proportion of actual normal drilling data and synthetic abnormal drilling data to achieve the drilling process state monitoring under small sample conditions. The specific steps are as follows:

[0081] (1) Pretreatment of drilling process data

[0082] During the drilling process, there are many types of drilling variables, and not all variables are closely related to the change of the drilling state. Through feature selection, features closely related to the drilling state can be screened out from more variables. Then, when generating abnormal drilling data, only the drilling process data that has a direct impact on the drilling state is generated, reducing the training volume of the data generation model and improving the data generation efficiency. Therefore, first, the chi-square test method is used to perform feature selection preprocessing on the drilling process data. Refer to Figure 4 , Figure 4It is a bar chart of the chi-square statistic corresponding to each feature. By querying the chi-square distribution critical value table, the critical value when the significance level is 0.1 and the degree of freedom is 1 is 2.706. This value is used as the threshold, as shown by the dashed line in the figure. Features corresponding to a statistic greater than 2.706 are selected as features with a relatively strong correlation with the drilling state, including features numbered 1, 2, 3, 4, 6, 7, 8, 9, 11, 13, 14, 15, 17, 18, 19, 20, 21, 22, 23, and 24.

[0083] (2)Training of the data generation model based on the improved generative adversarial network

[0084] Based on the selection of drilling process features, for the selected features, the data generation model is trained using the data of the corresponding features of two wells. During the model training process, different optimal model parameter combinations are designed for different types of drilling accidents as the initial values for parameter updates. The parameters are mainly the learning rate and the batch size. Refer to Figure 5 , Figure 5 For the generation effects of two types of drilling accident data when different parameter combinations are used as the initial values, the values shown in the squares in the figure represent the normalized maximum mean difference (MMD). The smaller the MMD, the better the generation effect. It can be seen from the figure that the optimal parameter combinations of the learning rate and the batch size corresponding to the slow-varying type accident and the sudden-change type accident are [0.0032, 8] and [0.0002, 16], respectively.

[0085] After selecting the optimal parameter combination as the initial value of the model training parameters, corresponding loss functions are designed for the generator G and the discriminator D respectively. Through the adversarial training of G and D , the model parameters are continuously updated until balance is achieved, and the model training is completed to obtain a data generation model with stable performance . Refer to Figure 6 , Figure 6 For the change trends of the losses of the generator and the discriminator with the number of iterations, at the beginning of the training, the loss of the discriminator is small, while the loss of the generator is large. After about 6000 iterations of training, G and D reach balance, and the data generation model is obtained.

[0086] (3) B Expansion of the sample data set of abnormal drilling states of wells

[0087] Using the data generation model and A to expand the sample data set of abnormal drilling states of wells B and Figure 7 ,Figure 7 This is a comparison between real accident data and synthetic accident data under stuck drill accidents. The rows in the figure represent the drilling process features processed by feature selection, and the columns represent the number of samples. It can be seen from the figure that the overall distribution is relatively similar.

[0088] Taking the torque data sequence of a drill stuck accident as an example, refer to Figure 8 , Figure 8 (a) is the torque series comparison between real data and synthetic data. It can be seen that the distribution of synthetic data is consistent with the change trend of real data. At the same time, the cumulative distribution function of torque is used as a quantitative evaluation indicator, such as Figure 8 As shown in (b), it can be seen that the distribution of synthetic data is the same as that of real data, which illustrates the effectiveness of data generation.

[0089] (4) Drilling process condition monitoring based on synthetic data

[0090] In realization B After the abnormal drilling data of the well is generated, the abnormal drilling status sample data set is expanded so that the ratio of normal drilling data to abnormal drilling data reaches 1:1, forming the drilling status monitoring model training set. On this basis, a drilling status monitoring model with PNN as the base classifier is established, with precision, accuracy, recall and As performance indicators, the monitoring effect is evaluated. Table 1 shows the monitoring effect when different data are used to establish the drilling status monitoring model. It can be seen from the table that the monitoring effect is the best when the generated data is used to establish the status monitoring model. Compared with the existing method of using other drilling data to establish the status monitoring model, all performance indicators have been improved.

[0091] Table 1 Monitoring effect when using different data to establish drilling status monitoring model (unit: %)

[0092]

[0093] In some embodiments, a drilling process small sample state monitoring device based on data enhancement is also provided, referring to Figure 9 , the device includes the following modules:

[0094] Preprocessing module 1 is used to process the reference well drilling data and target well drilling data Perform feature selection preprocessing to filter out features that are highly relevant to the drilling process;

[0095] Generative model training module 2 is used to improve the generative adversarial network by designing the model parameter configuration module and constructing the feature distance loss function according to the multi-time scale and multi-variable characteristics of the drilling process; based on the improved generative adversarial network, the preprocessed reference well abnormal drilling data is used And abnormal drilling data of the target well Train a data generation model ;

[0096] An abnormal drilling data generation module 3 for using the abnormal drilling data of the reference well As the input of the data generation model The output of the data generation model Is the synthetic abnormal drilling data of the target well ;

[0097] A drilling state monitoring module 4 for selecting a base classifier and combining the balanced drilling data set, including equal proportions of the synthetic abnormal drilling data of the target well And actual normal drilling data , Establish a drilling state monitoring model to realize the drilling process state monitoring under the condition of small samples.

[0098] The beneficial effects of the present invention are: through the improved generative adversarial network, the sample data set under abnormal drilling conditions is expanded, the generalization ability of the drilling state monitoring model is improved, the monitoring effect is improved, and safety guarantee is provided for the drilling process, which has practicability and applicability.

[0099] It should be noted that in this article, the terms "including", "comprising" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or elements inherent to such process, method, article or system. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, article or system including that element.

[0100] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. Among the several device unit claims, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words can be interpreted as identifiers.

[0101] The above is only the preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be included in the patent protection scope of the present invention by the same token.

Claims

1. A small-sample condition monitoring method for the drilling process based on data augmentation, characterized in that: including the following steps: S1: Preprocess the drilling data of the reference well and the drilling data of the target well by feature selection, and screen out the features that are highly correlated with the drilling process; S2: Considering the multi-time scale and multi-variable characteristics of the drilling process, improve the generative adversarial network by designing a model parameter configuration module and constructing a feature distance loss function; based on the improved generative adversarial network, use the pre-processed reference well abnormal drilling data in step S1 and the target well abnormal drilling data to train and obtain a data generation model ; Among them, the steps for improving the generative adversarial network are specifically: S2.2: For mutation types and slow change types of drilling accidents, set different initial values for the training parameters of the data generation model, including the learning rate lr and batch size bs : In the formula, is the label of the drilling accident category, and are the parameter combinations corresponding to the mutation type and the slow change type respectively; S2.3: Update the network parameters by means of stochastic gradient descent; calculate the magnitude and direction of the gradient descent according to the loss function, and the generator loss function and the discriminator loss function are respectively: Wherein, is the synthetic data sequence corresponding to the f th feature, is the real data sequence corresponding to the f th feature, r is the number of features after feature selection, is the loss between the two sequences, m is the number of samples participating in network training; S3: Use the abnormal drilling data of the reference well as the input of the data generation model . The output of the data generation model is the synthetic abnormal drilling data of the target well ; S4: Select a base classifier, and combine the balanced drilling dataset, including equally proportioned synthetic abnormal drilling data of target wells and actual normal drilling data , to establish a drilling state monitoring model and achieve drilling process state monitoring under small sample conditions.

2. The small-sample condition monitoring method for the drilling process based on data augmentation according to claim 1, wherein: Step S1 includes: Using the chi-square test method, for the drilling data of the reference well and the drilling data of the target well perform feature selection preprocessing to screen out features with a chi-square statistic greater than the probability threshold.

3. The small-sample condition monitoring method for the drilling process based on data augmentation according to claim 2, wherein: The chi-square test method is used to perform feature selection preprocessing on the drilling data of the reference well and the drilling data of the target well The steps of screening out features with a chi-square statistic greater than the probability threshold include: S1.1: Calculate the drilling data of the reference well and the drilling data of the target well for the chi-square statistic of each feature , and the calculation formula is: In the formula, is the observed frequency of the actual data, is the expected frequency, K is the product of the number of test hypotheses and the number of drilling state categories, k = 1, 2, ..., K ; S1.2: Compare the chi-square statistics of each feature with the probability threshold T in terms of magnitude. If is greater than T , retain the corresponding feature; otherwise, remove the corresponding feature.

4. The small-sample condition monitoring method for the drilling process based on data augmentation according to claim 1, wherein: Step S1 includes: Using the Xgboost method, for the drilling data of the reference well and the drilling data of the target well Perform feature selection preprocessing to screen out features that are highly relevant to the drilling process.

5. The small sample condition monitoring method for the drilling process based on data augmentation according to claim 1, characterized in that: In step S2, the construction steps of the generative adversarial network are: S2.1: Construct a generative adversarial network. The generator G is used to generate synthetic drilling data, and the discriminator D is used to distinguish synthetic data from real data. By minimizing the generator loss and maximizing the discriminator loss, the network training is completed, G and D when reaching equilibrium, G it is the data generation model obtained through training ; Given m abnormal drilling samples of the target well , and m abnormal drilling samples of the reference well , taking the samples as the input of the generator G , and the samples and the generator output together as the input of the discriminator D , then the objective function is expressed as: Wherein, is the data distribution of the target well, is the data distribution of the reference well, is the generator output, is the discriminator output, is the discriminator loss, is the generator loss.

6. The small-sample condition monitoring method for the drilling process based on data augmentation according to claim 1, characterized in that: In step S4, the base classifier includes at least one of a probabilistic neural network, an artificial neural network, and a convolutional neural network.

7. A small-sample condition monitoring device for the drilling process based on data augmentation, characterized in that: Steps for implementing the small-sample state monitoring method for the drilling process based on data augmentation according to any one of claims 1 to 6.

Images