Early overflow intelligent monitoring method and system based on multivariate time series

Abstract

The invention discloses a multivariate time series-based early overflow intelligent monitoring method and system, and relates to the technical field of petroleum drilling engineering, and the method comprises the steps: obtaining a time sequence sample data set of sample drilling, and obtaining a time sequence feature sample data set corresponding to the time sequence sample data set; amplifying the original overflow data in the overflow data set to a preset number through a preset initial model to obtain an amplified overflow data set; inputting the non-overflow data set and the amplified overflow data set into an initial model, and training the initial model to obtain a trained prediction model corresponding to the initial model; the test data set of the to-be-tested drilling well is obtained and input into the prediction model to be processed, the prediction result corresponding to the to-be-tested drilling well is obtained, the prediction result comprises an overflow event and a non-overflow event, overflow can be recognized as early as possible and processed in time, blowout accidents are avoided, and the safety of the to-be-tested drilling well is improved. And irreparable property loss and casualties are prevented.

Description

Technical Field

[0001] The present invention relates to the technical field of oil drilling engineering, and more specifically, to an early overflow intelligent monitoring method and system based on multivariate time series. Background Art

[0002] In oil drilling projects, overflow is a common complex underground accident. If the overflow is not identified and handled in time, it may develop into a blowout, causing irreparable property losses and casualties.

[0003] There are many reasons for overflow. The most fundamental reason is that the pressure in the well loses balance, the bottom hole pressure is lower than the formation pressure, the formation gas or fluid flows into the wellbore, the drilling fluid outlet flow is greater than the inlet flow, and the fluid in the wellbore automatically overflows from the wellhead after the well is shut down. The main reasons for overflow are: (1) Inaccurate prediction of formation pressure: In recent years, although the ability to predict formation pressure has been greatly improved, since the geological correlation mainly depends on the interpretation, it is possible to encounter unexpected abnormally high-pressure layers at any time during the entire drilling process; (2) The drilling fluid level in the wellbore drops: No drilling fluid is injected when drilling is started or the volume of drilling fluid injected is less than the volume of the drill bit pulled out, and caves or cracks are encountered during drilling, resulting in well leakage. The well leakage is not discovered in time or the drilling fluid is not replenished after the drilling fluid leaks, or the drilling fluid is insufficiently replenished, resulting in a drop in the drilling fluid level in the wellbore; (3) The density of the drilling fluid decreases: In general, the bottom hole pressure is higher than the formation pore pressure supporting the wellbore. The design of drilling mud density should follow the two key boundaries of formation pressure and fracture pressure (pressure window). However, due to the particularity and uncertainty of geology, it is almost impossible to accurately predict pore pressure all the time. Then, when the pore pressure prediction is inaccurate, the bottom hole pressure loses balance, the intrusion of formation fluid (oil, gas, water) causes the drilling fluid density to decrease or the drilling fluid weight is lower than expected, the drilling fluid density suddenly drops, causing the wellbore pressure to decrease, or the formation pressure is higher than expected, so the static pressure exerted by the specific gravity of the drilling fluid is not enough to balance the formation pressure in the wellbore; (4) The pumping effect during drilling causes the wellbore liquid column pressure to drop: when starting, the drill bit moves upward too fast, resulting in a pumping effect, which causes the liquid column pressure to decrease, resulting in formation pressure greater than the bottom hole pressure; (5) The annular pressure loss acting on the bottom hole disappears when the circulation is stopped: when the pump is turned on to circulate the drilling fluid, the pump pressure is mainly consumed by the friction of the circulation system and the pressure drop at the drill bit water hole. When the pump is turned on to drill, the formation pressure is balanced by the static liquid column pressure and the annular friction. When the pump is stopped, the disappearance of the annular pressure loss will cause the bottom hole pressure to be unbalanced, resulting in overflow.

[0004] During the drilling process, overflow events are accidental. From the perspective of machine learning, the essence of early overflow monitoring is a binary classification task. Samples with overflow are called positive samples, and samples without overflow are called negative samples. However, the number of overflow event samples is much smaller than the number of samples without overflow. The overflow data set is extremely unbalanced. Therefore, early overflow monitoring can be regarded as a binary classification problem using unbalanced small sample data. The early overflow intelligent monitoring method based on machine learning has a great imbalance between positive and negative sample classes, which makes it difficult to train the model because the overflow samples are far less than the normal samples. In addition, the early Most of the algorithms are designed based on balanced data sets, and do not consider the problem of data imbalance. The monitoring effect is relatively unsatisfactory. Even before this, some people have introduced generative adversarial networks into early overflow monitoring and constructed an early overflow intelligent monitoring method based on cost-sensitive learning. However, the extraction module for the wellhead time series feature parameters in this method is relatively simple, and it is unable to extract the deep-level information implied in the complex time series features. In addition, the early overflow intelligent monitoring method based on cost-sensitive learning finally uses ensemble learning as a classifier. The process of using ensemble learning as a classifier is relatively complex and cumbersome, and it is difficult to apply it in practice. Summary of the invention

[0005] The purpose of the present invention is to provide an early overflow intelligent monitoring method and system based on multivariate time series, which provides a new early overflow monitoring method for machine learning due to the scarcity of sample data such as overflow monitoring during drilling, and can achieve early identification and timely processing of overflow, avoid blowout accidents, and prevent irreparable property losses and casualties.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions:

[0007] In a first aspect, an embodiment of the present application provides an early overflow intelligent monitoring method based on a multivariate time series, comprising the following steps:

[0008] Acquire a time series sample data set of sample drilling, and perform preprocessing and feature extraction on the time series sample data set to obtain a time series feature sample data set corresponding to the time series sample data set, wherein the time series feature sample data set includes an overflow data set and a non-overflow data set;

[0009] The original overflow data in the overflow data set is amplified to a preset amount through a preset initial model to obtain an amplified overflow data set;

[0010] Inputting the non-overflow data set and the expanded overflow data set into the initial model, and training the initial model to obtain a prediction model that corresponds to the initial model and has been trained;

[0011] A test data set of the well to be tested is obtained, and the test data set is input into the prediction model for processing to obtain the prediction result corresponding to the well to be tested, and the prediction result includes overflow events and non-overflow events.

[0012] The beneficial effects of the present invention are as follows: by preprocessing and feature extracting the acquired drilling timing parameters, and based on the drilling timing parameters after the preprocessing and feature extraction, the parameters of the drilling timing parameters that will cause overflow events are amplified, thereby effectively solving the problem of serious imbalance between the parameters for overflow events and the parameters for overflow events, as well as the problem of too few parameters for overflow time leading to low classification accuracy and poor generalization ability of the prediction model, and using the parameters for training the initial model to balance the parameters occurring at positive and negative times, ultimately achieving a better initial model training and making the obtained prediction model more accurate.

[0013] The real drilling time series data set is extracted using the LSTM network, and then the output vectors of each part are fused into a feature vector; secondly, the problems of low model classification accuracy and poor generalization ability caused by too few overflow samples are addressed.

[0014] Based on the above technical solution, the present invention can also be improved as follows.

[0015] Further, the above initial model includes a feature extraction module, and the feature extraction module includes a data preprocessing unit and a feature extraction unit; wherein, the time series sample data set is preprocessed and feature extracted to obtain a time series feature sample data set corresponding to the time series sample data set, and the time series feature sample data set includes an overflow data set and a non-overflow data set, including:

[0016] The data preprocessing unit sequentially performs data cleaning, data noise reduction, missing value processing and normalization processing on the sample data in the time series sample data set to obtain a feature data set to be extracted corresponding to the time series sample data set;

[0017] Through multiple LSTM networks in the feature extraction unit, the multivariate time series features of different time ranges in the time series sample data set are captured respectively, and multiple time series feature sets of different time ranges are obtained;

[0018] The multiple time series feature sets are subjected to feature fusion, and a data set obtained after the feature fusion is determined as a time series feature sample data set.

[0019] The beneficial effects of adopting the above further scheme are: it can effectively reduce the dimension of data in the time series sample data set, improve the robustness to noise, improve data quality, enhance the correlation between data features and overflow events, reduce the difficulty of identifying the initial model and improve prediction accuracy.

[0020] Furthermore, the above initial model also includes a sample set amplification module, and the sample set amplification module includes a generator G and a generator F; wherein, the original overflow data in the overflow data set is amplified to a preset amount through the preset initial model to obtain an amplified overflow data set, including:

[0021] The original non-overflow data in the non-overflow data set is input into the generator G for processing to generate augmented data;

[0022] The amplified data and the original overflow data in the overflow data set are input into the generator F for processing respectively, so as to generate overflow identification data corresponding to the amplified data and overflow data to be tested corresponding to the original overflow data;

[0023] The overflow identification data is compared with the overflow data to be tested. If the difference between the overflow identification data and the overflow data to be tested is within a preset difference range, the overflow data to be tested is determined to be amplified overflow data. The amplified overflow data set includes the original overflow data and the amplified overflow data.

[0024] The beneficial effect of adopting the above further scheme is: the data generated by generator G is input into generator F for further processing, and the data generated twice by generator F are compared and judged, thereby increasing the difference of samples while maintaining the overall outline of the samples.

[0025] Furthermore, the sample set amplification module further includes a discriminator Dx and a discriminator Dy, and the method further includes:

[0026] The generator G generates a first function loss value of the amplified data based on the original non-overflow data by calculating the discriminator Dy, and if the first function loss value is within a first preset range, the amplified data with the first function loss value within the first preset range is input into the generator F for processing;

[0027] The generator F calculates the second function loss value of the overflow data to be tested based on the original overflow data through the discriminator Dx. If the second function loss value is within the second preset range, the overflow data to be tested with the second function loss value within the second preset range is compared with the overflow identification data.

[0028] Furthermore, the above method also includes:

[0029] The total function loss value of the sample set expansion module is calculated according to the first function loss value and the second function loss value;

[0030] If the total function loss value meets the preset amplification conditions, the original overflow data in the overflow data set is amplified to the preset number through the sample set amplification module that meets the amplification conditions. If the total function loss value does not meet the amplification conditions, the module parameters of the sample set amplification module are adjusted, and the total function loss value is recalculated based on the adjusted module parameters until the total function loss value meets the amplification conditions.

[0031] Furthermore, the above total function loss value is expressed by the first formula, which is:

[0032] L(G,F,D X ,D Y )=L GAN (G,D Y ,N N ,N P )+L GAN (F,D X ,N P ,N N )+L cGAN (x,y,z)+λL cyc (G,F)+L1(x,y c );

[0033] In the formula, L(G,F,D X ,D Y ) represents the total function loss value, L GAN (G,D Y ,N N ,N P ) represents the loss value of the first function, L GAN (F,D X ,N P ,N N ) represents the loss value of the second function, L cGAN (x,y,z) represents the noise loss value, λL cyc (G, F) represents the cycle consistency loss value, L1(x,y c ) represents the difference between the overflow identification data and the overflow data to be tested, and λ represents an adjustable parameter.

[0034] Furthermore, the initial module further includes a classification module, wherein the non-overflow data set and the amplified overflow data set are input into the initial model, and the initial model is trained to obtain a prediction model corresponding to the initial model and trained, including:

[0035] Inputting the non-overflow data set and the amplified overflow data set into the classification module for processing respectively, obtaining a first processing result corresponding to the non-overflow data set and a second processing result corresponding to the amplified overflow data set;

[0036] Whether the initial model meets the preset training end condition is determined according to the first processing result and the second processing result. If the initial model meets the training end condition, the initial model that meets the training end condition is determined as the prediction model. If the initial model does not meet the training end condition, the model parameters of the initial model are adjusted, and the initial model is continued to be trained based on the adjusted model parameters until the first processing result and the second processing result meet the training end condition.

[0037] In a second aspect, an embodiment of the present application provides an early overflow intelligent monitoring system based on a multivariate time series, which is applied to an early overflow intelligent monitoring method based on a multivariate time series in any one of the first aspects, including:

[0038] The sample acquisition processing module is used to acquire a time series sample data set of sample drilling, and perform preprocessing and feature extraction processing on the time series sample data set to obtain a time series feature sample data set corresponding to the time series sample data set, wherein the time series feature sample data set includes an overflow data set and a non-overflow data set;

[0039] A sample amplification module, used to amplify the original overflow data in the overflow data set to a preset amount through a preset initial model to obtain an amplified overflow data set;

[0040] A model training module is used to input the non-overflow data set and the amplified overflow data set into the initial model, and train the initial model to obtain a prediction model that corresponds to the initial model and has been trained;

[0041] The model prediction processing module is used to obtain the test data set of the well to be tested, and input the test data set into the prediction model for processing to obtain the prediction results corresponding to the well to be tested, and the prediction results include overflow events and non-overflow events.

[0042] In a third aspect, an embodiment of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any one of the methods in the first aspect when executing the computer program.

[0043] In a fourth aspect, an embodiment of the present application provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable a computer to execute any one of the methods in the first aspect.

[0044] Compared with the prior art, the present invention has at least the following beneficial effects:

[0045] Firstly, the preprocessed drilling time series data set is feature extracted, and then the output vectors of each part are fused into a feature vector; secondly, in order to solve the problems of low model classification accuracy and poor generalization ability caused by too few overflow samples, the overflow samples are amplified through a cycle consistency generative adversarial network, and then the amplified sample data set is spliced ​​with the original sample data set and input into a multi-layer perceptron for classification; the relationship between the change of characteristic parameters and early overflow is analyzed by an intelligent model, so as to achieve the purpose of intelligent monitoring of early overflow.

[0046] The data generated by the forward generator is input into the reverse generator, which increases the diversity of the samples while maintaining the overall outline of the samples.

[0047] The classification accuracy of the method of the present application is compared with that of the undersampling method, the oversampling method and the mixed sampling method. The results show that the classification performance of these three sampling methods for negative samples is higher than their classification performance for positive samples, and the overall classification effect of the method of the present invention is significantly better than that of the three sampling methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0049] Figure 1 It is a framework diagram of the prediction model in the embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the structure of an LSTM network in an embodiment of the present invention;

[0051] Figure 3 This is a flow chart of feature extraction by LSTM network in an embodiment of the present invention;

[0052] Figure 4 Schematic diagram of the process of the sample set amplification module in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of the structure of a classification module in an embodiment of the present invention;

[0054] Figure 6 This is a comparison chart of classification performance of models trained by different methods in an embodiment of the present invention;

[0055] Figure 7 is a method flow chart of the monitoring method in an embodiment of the present invention;

[0056] Figure 8 A connection diagram of a monitoring system in an embodiment of the present invention;

[0057] Fig. 9Schematic diagram of the connection of electronic equipment in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0059] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0061] In the description of the embodiments of the present invention, "plurality" means at least 2.

[0062] Example 1

[0063] The present application embodiment provides an early overflow intelligent monitoring method based on multivariate time series, see Figure 7 , including the following steps:

[0064] S1, obtaining a time series sample data set of sample drilling, and performing preprocessing and feature extraction on the time series sample data set to obtain a time series feature sample data set corresponding to the time series sample data set, wherein the time series feature sample data set includes an overflow data set and a non-overflow data set.

[0065] Among them, the time series sample data set may include data of multiple sample drillings. The time series sample data set includes multiple parameters of the corresponding sample drillings arranged in time series. The multiple parameters may include riser pressure difference, mud pool and total pool volume difference, inlet and outlet drilling fluid flow difference, inlet and outlet drilling fluid density difference, inlet and outlet drilling fluid temperature difference, inlet and outlet drilling fluid conductivity difference, drilling time, that is, the time series sample data set contains multiple corresponding drilling parameter conditions arranged in chronological order; specifically, among these parameters, some parameters will cause drilling overflow when corresponding to the situation, so this part of the parameters is used as the overflow data set. In addition, another part of the parameters is normal and will not cause drilling overflow, so the part of the parameters under the right situation is determined as the non-overflow data set.

[0066] Optionally, the initial model includes a feature extraction module, and the feature extraction module includes a data preprocessing unit and a feature extraction unit; wherein the time series sample data set is preprocessed and feature extracted to obtain a time series feature sample data set corresponding to the time series sample data set, and the time series feature sample data set includes an overflow data set and a non-overflow data set, including:

[0067] The data preprocessing unit sequentially performs data cleaning, data noise reduction, missing value processing and normalization processing on the sample data in the time series sample data set to obtain a feature data set to be extracted corresponding to the time series sample data set.

[0068] Among them, when processing the time series sample data set, data preprocessing and feature extraction processing can be performed in sequence; specifically, the purpose of data preprocessing is to effectively reduce the dimension of the data in the time series sample data set, improve the robustness to noise, and improve data quality, enhance the correlation between data features and overflow events, reduce the recognition difficulty of the initial model and improve prediction accuracy; among them, preprocessing can include data cleaning, data denoising, normalization, removal of noise data and missing values, etc., and min-max Normalization can be used to normalize the data after removing noise data and missing values. The normalization function is as follows:

[0069]

[0070] In the formula, x represents the original data in the time series sample data set; Max(x) represents the maximum value of a parameter in the time series sample data set; Min(x) represents the minimum value of the parameter in the time series sample data set; x' represents the data after the original data in the time series sample data set is standardized.

[0071] Through multiple LSTM networks in the feature extraction unit, the multivariate time series features of different time ranges in the time series sample data set are captured respectively, and multiple time series feature sets of different time ranges are obtained.

[0072] Among them, when extracting features from the above data, multiple LSTM networks can be used, such as Figure 1 and Figure 2 As shown, four LSTM networks can be used for feature extraction. The four LSTM networks extract features of the drilling parameters obtained during the drilling process, and the time steps of the sliding windows of the four LSTM networks in the feature extraction process can be set to 3, 5, 7, and 9, respectively, to capture multivariate time series features in different ranges. Therefore, the time series feature data sets N1, N2, N3, and N4 corresponding to four different time steps can be obtained respectively.

[0073] Specifically, the LSTM network contains four interaction layers, namely three gate layers and one tanh layer; the first interaction layer is called the "forget gate layer", which is responsible for determining the information that needs to be forgotten in the current step; the output ft of the forget gate is a number ranging from 0 to 1, indicating the probability of forgetting the relevant information; the output of the forget gate is multiplied by the processor state to determine which parts of the processor state in the previous step should be retained and passed to the next step; the calculation formula of the forget gate is as follows:

[0074] i t =σ[W i (h t-1 ,x t )+b i ].

[0075] Specifically, the output i of the input gate layer t Determines which parts of the candidate values ​​calculated by the tanh layer will be retained and used to update the processor state; the third interactive layer is the tanh layer, which is responsible for creating a new candidate value that can be added to the processor state to update its information. The result of the tanh layer is calculated by the following formula: C t '=tanh[W c (h t-1 ,x t )+b c ].

[0076] Specifically, after three layers of calculation, the old processor state C (t-1) Carrying memory information and new information candidate value C t The specific method is as follows: C t =f t C t-1 +i t C t '; The calculation result of the forget gate f t Decided to start from C (t-1) What information is forgotten in the input gate? t Decide on C t What information to keep and add.

[0077] Specifically, the last layer of the LSTM network is the “output gate layer”, which uses the updated processor state to generate the output value of the LSTM as shown in the following formula: t =σ[W o (h t-1 ,x t )+b o ]tanhC t .

[0078] Specifically, the feature extraction process is as follows: Figure 1 and Figure 3 As shown in the figure, the LSTM network has the ability to capture the temporal dependencies in time series data based on the memory unit's ability to maintain memory. In order to fully tap the high-dimensional, multi-scale features and deep-level features implicit in the complex and multivariate time series parameters, four LSTM networks can be constructed to extract features of the drilling parameters obtained during the drilling process. Then, the time series feature data sets N1, N2, N3 and N4 corresponding to four different time steps can be obtained respectively.

[0079] The multiple time series feature sets are subjected to feature fusion, and a data set obtained after the feature fusion is determined as a time series feature sample data set.

[0080] Among them, the four feature data sets (N1, N2, N3 and N4) can be fused and added and averaged to obtain the final time series feature sample data set M(b). The time series feature sample data set M(b) contains the overflow data set N P and non-overflow dataset N N .

[0081] S2, amplifying the original overflow data in the overflow data set to a preset amount through a preset initial model to obtain an amplified overflow data set.

[0082] Among them, due to the accumulation of drilling experience and other factors, overflow rarely occurs during the drilling process. Therefore, in the time series feature sample data set, the number of overflow data sets is much smaller than the number of non-overflow data sets. In the time series feature sample data set formed in this way, the data when the positive and negative events occur will be seriously unbalanced, and the training of the initial model through such data sets cannot achieve the purpose, but will lead to problems such as low model classification accuracy and poor generalization ability; therefore, before training the model, it is necessary to balance the two types of data at positive and negative times in the time series feature sample data set. Since the number of overflow data sets in the time series feature sample data set is small, it is necessary to expand the number of overflow data sets so that the number of overflow data sets is the same or approximately the same as the number of non-overflow data sets, so as to achieve the purpose of balance; specifically, when carrying out the specific implementation, it is necessary to determine the amount of expansion based on the number of non-overflow data sets, so after the number of non-overflow data sets is determined, a preset number can be set, and the data of the overflow data set can be expanded according to the preset number.

[0083] Optionally, the initial model further includes a sample set amplification module, and the sample set amplification module includes a generator G and a generator F; wherein the original overflow data in the overflow data set is amplified to a preset amount through a preset initial model to obtain an amplified overflow data set, including:

[0084] The original non-overflow data in the non-overflow data set is input into the generator G for processing to generate augmented data.

[0085] The amplified data and the original overflow data in the overflow data set are input into the generator F for processing respectively, so as to generate overflow identification data corresponding to the amplified data and overflow data to be tested corresponding to the original overflow data.

[0086] Among them, the sample set expansion module can supplement overflow samples on the CycleGAN structure through the cycle consistency generative adversarial network (CycleGAN, CycleGenerative Adversarial Network), optimize the module by changing the loss function, and input the data generated by the forward generator into the reverse generator, thereby maintaining the overall outline of the sample while increasing the diversity of the sample.

[0087] Specifically, the overall processing flow in the sample set amplification module is as follows: Figure 1 and Figure 4 As shown, from the non-overflow data set N N Take a data x (original non-overflow data) from the (non-overflow data set) and use the generator G to generate data y′ (augmented data); similarly, take a data x (original non-overflow data) from the overflow data set N P Take data y (original overflow data) from (overflow data set) and input it into generator F to generate data x' (overflow data to be tested); at the same time, input data y' (amplified data) into generator F to generate data x" (overflow identification data).

[0088] The overflow identification data is compared with the overflow data to be tested. If the difference between the overflow identification data and the overflow data to be tested is within a preset difference range, the overflow data to be tested is determined to be amplified overflow data. The amplified overflow data set includes the original overflow data and the amplified overflow data.

[0089] Among them, the data x" (overflow identification data) and x' (overflow data to be tested) are both generated by the generator F, and the difference between the data x" (overflow identification data) and the data x' (overflow data to be tested) is used to normalize the generated data, that is, the difference between the two cannot exceed the preset difference range; specifically, according to the above steps, a preset number of amplified overflow data can be generated, and the amplified overflow data is saved in the amplified overflow data set, so the amplified overflow data set includes the original overflow data and the amplified overflow data, so that the sum of the number of original overflow data and the number of amplified overflow data is balanced with the number of original non-overflow data.

[0090] Optionally, the sample set amplification module further includes a discriminator Dx and a discriminator Dy, and the method further includes:

[0091] The generator G generates the first function loss value of the amplified data based on the original non-overflow data through the discriminator Dy. If the first function loss value is within the first preset range, the amplified data with the first function loss value within the first preset range is input into the generator F for processing.

[0092] Among them, the loss value of the first function can be expressed by the following formula:

[0093]

[0094] In the formula, G is the generator G, Dy is the discriminator Dy, N N represents the non-overflow data set, N P represents the overflow data set, y is the original overflow data, G(y) is the amplified data generated by generator G, F(G(y)) is the overflow identification data generated by generator F, represents the loss value of the discriminator Dy; Represents the loss value of the generator G.

[0095] The generator F calculates the second function loss value of the overflow data to be tested based on the original overflow data through the discriminator Dx. If the second function loss value is within the second preset range, the overflow data to be tested with the second function loss value within the second preset range is compared with the overflow identification data.

[0096] Among them, the expression of the second function loss value corresponds to the expression of the first function loss value, and the second function loss value is expressed as:

[0097]

[0098] In the formula, F is the generator F, Dx is the discriminator Dx, N N represents the non-overflow data set, N P represents the overflow data set, x represents the original non-overflow data, F(x) is the overflow data to be tested generated by generator F, G(F(x)) represents the data generated by the overflow data to be tested through generator G, Represents the loss value of the discriminator Dx; Represents the loss value of the generator F.

[0099] Specifically, the above-mentioned discriminator is used to distinguish between true and false. When the discriminator (discriminator Dx and discriminator Dy) discriminates the input data, the output result is 1, indicating that the discriminated input data is data in the original data (original overflow data and original non-overflow data), and 0 indicates that the discriminated input data is generated data (overflow data to be tested, amplified data or overflow identification data); because the discriminator may make misjudgments, that is, when the training level is good, the output of the data generated by the generator after inputting into the discriminator is 0.5, then the discriminator identifies that the data has a half probability of being true and a half probability of being false; by training the discriminator to identify real samples and generated samples, its ability to identify true and false is improved, and the generator continuously generates samples of better quality; the discriminator and the generator confront and learn from each other, and finally reach a balance, that is, the generator can no longer generate better samples, and at this time the probability that the discriminator identifies the generated sample as true is 0.5.

[0100] Optionally, the above method further includes:

[0101] The total function loss value of the sample set expansion module is calculated according to the first function loss value and the second function loss value;

[0102] If the total function loss value meets the preset amplification conditions, the original overflow data in the overflow data set is amplified to the preset number through the sample set amplification module that meets the amplification conditions. If the total function loss value does not meet the amplification conditions, the module parameters of the sample set amplification module are adjusted, and the total function loss value is recalculated based on the adjusted module parameters until the total function loss value meets the amplification conditions.

[0103] Optionally, the total function loss value is expressed by a first formula, which is:

[0104] L(G,F,D X ,D Y )=L GAN (G,D Y ,N N ,N P )+L GAN (F,D X ,N P ,N N )+L cGAN (x,y,z)+λL cyc (G,F)+L1(x,y c );

[0105] In the formula, L(G,F,D X ,D Y ) represents the total function loss value, L GAN (G,D Y ,N N ,N P ) represents the loss value of the first function, LGAN (F,D X ,N P ,N N ) represents the loss value of the second function, L cGAN (x,y,z) represents the noise loss value, λL cyc (G, F) represents the cycle consistency loss value, L1(x,y c ) represents the difference between the overflow identification data and the overflow data to be tested, λ represents an adjustable parameter; Specifically, in the first formula, x represents the original non-overflow data, y represents the original overflow data, z is the random vector noise, G represents the generator G, F represents the generator F, Dx represents the discriminator Dx, Dy represents the discriminator Dy, N N represents the non-overflow data set, N P represents the overflow dataset, y c Indicates the overflow data to be tested.

[0106] Specifically, in the first formula, L cyc (G, F) is the cycle consistency loss, which is specifically expanded as follows:

[0107]

[0108] In the formula, F(G(x)) means that the data first passes through the generator G, and the generated data is used as the input of the generator F, so that a non-overflow dataset N can be generated. N N N ',N N 'Used to measure with N N The gap between the generated distribution and the original data is whether the generated distribution is consistent with the distribution of the original data.

[0109] Among them, in the sample set expansion module, each time data is generated, it must be generated in reverse. The difference between x" and the original data is calculated by regenerating the data x", so that the generated data is close to the original data and the data quality is improved, but the data diversity will be constrained; through the data reconstruction step, the contour is generated and then the data is generated. The contour is used as an intermediary to expand the difference with the original data by generating data twice, which is conducive to generating different overflow data.

[0110] Specifically, in the first formula above, L cGAN (x,y,z) is represented as:

[0111] L cGAN (x,y,z)=E x,y [lnD(x,y)]+E x,z [ln(1-D(x,G(x,z))];

[0112] In the formula, z represents the random noise vector; x represents the original non-overflow data; y represents the original overflow data; E x,y [lnD(x,y)] represents the discriminator loss value; E x,z [ln(1-D(x,G(x,z))] represents the generator loss value.

[0113] S3, inputting the non-overflow data set and the amplified overflow data set into the initial model, and training the initial model to obtain a prediction model that corresponds to the initial model and has been trained.

[0114] Among them, through the various steps in the above S2, an expanded overflow data set with a balanced number of data in the non-overflow data set can be obtained, and the expanded overflow data set includes the original overflow data and the expanded overflow data; thereby improving the accuracy and generalization ability of the trained model.

[0115] Optionally, the initial module further includes a classification module; wherein the non-overflow data set and the amplified overflow data set are input into the initial model, and the initial model is trained to obtain a prediction model corresponding to the initial model and trained, including:

[0116] The non-overflow data set and the amplified overflow data set are respectively input into the classification module for processing to obtain a first processing result corresponding to the non-overflow data set and a second processing result corresponding to the amplified overflow data set.

[0117] Whether the initial model meets the preset training end condition is determined according to the first processing result and the second processing result. If the initial model meets the training end condition, the initial model that meets the training end condition is determined as the prediction model. If the initial model does not meet the training end condition, the model parameters of the initial model are adjusted, and the initial model is continued to be trained based on the adjusted model parameters until the first processing result and the second processing result meet the training end condition.

[0118] Among them, the classification module can use a multilayer perceptron (MLP) for the final classification and complete the binary classification task between overflow and non-overflow events; the classification module has a simple structure, a controllable number of input variables, and strong operability, and can be well applied to overflow and non-overflow classification; specifically, the MLP consists of an input layer, a hidden layer, and an output layer. The input layer is used to receive data from the outside and provide data to the neural network, and the hidden layer is used for network optimization. The number of neurons is equal to the input and output attributes in these layers. The network structure of the MLP network in the binary classification case is as follows Figure 1 and Figure 5As shown in the figure, a multilayer perceptron is a mathematical function that maps a set of input values ​​to output values. The function is composed of multiple simple activation functions. The application of these mathematical functions layer by layer is considered to provide a new representation for the input data. Its purpose is to approximate a certain function f*. In the classification problem, y=f*(x) maps the input data x to the category y. The network can be defined as a mapping y=f(x;w). By learning the value of the weight parameter w, the best approximate function is obtained.

[0119] Specifically, MLP combines two or more functions layer by layer to form a chain network structure. (1) 、f (2) The connection with the three functions of g forms f(x)=g(f (2) (f (1) (x))) is the most common network structure of the feedforward neural network, where f (1) With f (2) They are called the first and second layers of the network respectively, and the length of the chain is called the depth of the network model; g is called the output layer of the network. During the training process of the network model, the network model f(x) is made to approach the value of the real model f*(x) as much as possible. Usually, f (1) With f (2) In the entire network, it plays the role of filtering and extracting features from the data, while g and softmax functions transform the extracted features into output dimensions and determine which category they belong to; f (1) With f (2) As a function, it is generally called an activation function, which is used to activate or inhibit the state of a single neuron.

[0120] Specifically, the classification process of the classification module concatenates all the data sets obtained by the feature extraction module and the data set expansion module to obtain the final data set for classification as the input of the classification model, and the feature parameters input by the input layer are classified and processed by the MLP network to obtain the final classification results (the first processing result and the second processing result), and then the first processing result and the second processing result are compared with the actual results to obtain the correct processing result of the first processing result and the second processing result. In the process of continuous classification, training, and learning adjustment, the probability of the correct processing result of the classification module will become larger and larger, among which the probability of the correct processing result is greater than 95%, which can be used as the end condition of training, so that the initial model composed of the feature extraction module, the data set expansion module and the classification module is determined as the prediction model finally put into use.

[0121] S4, obtaining a test data set of the well to be tested, and inputting the test data set into a prediction model for processing to obtain a prediction result corresponding to the well to be tested, the prediction result including overflow events and non-overflow events.

[0122] The output of the prediction model is overflow or normal operating conditions (overflow events and non-overflow events), so as to achieve the purpose of predicting overflow events or non-overflow events in the well according to the parameters of the well (the test data set of the well to be tested).

[0123] Specifically, this embodiment selects the actual drilling data of the Penglai gas field in the Sichuan Basin in western China as the original data, and records the cumulative changes of each characteristic parameter within a certain period of time (60s). In order to simulate the development of the block in different periods and test the generalization ability of the model, data sets of various sizes are constructed; first, the following test data set A is constructed for verifying the model, and the specific information is shown in Table 1 below.

[0124] Table 1 Dataset A

[0125]

[0126] Among them, the dataset includes a total of 1200 sample groups, including 96 positive and 1104 negative sample groups respectively, and the sample imbalance rate is as high as 11.5, which is a typical unbalanced dataset.

[0127] Specifically, for the binary classification problem, in order to comprehensively evaluate the model used for imbalanced data classification and indirectly compare the quality of the binary classification model, this embodiment combines the actual sample labels and the predicted labels in pairs, and there are four possible results. The results of these four combinations are also called confusion matrices, as shown in Table 2.

[0128] Table 2 Confusion matrix

[0129]

[0130] In predictive classification models, the higher the accuracy, the better. That is to say, in the confusion matrix, the best result is to have a large number of TP and TN, and a small number of FP and FN. However, the confusion matrix is ​​a statistical number, and sometimes it is difficult to define the quality of the model by simply counting the number. Therefore, in order to better evaluate the quality of the model, the confusion matrix introduces the following indicators based on the basic statistical results: Accuracy, Recall, Precision and F-Measure, where:

[0131]

[0132]

[0133] Among them, Accuracy represents the accuracy rate, which is the proportion of all correct results of the classification model to the total value for the entire model; Recall represents the recall rate, which is the proportion of correct model predictions among all results whose true values ​​are positive; Precision represents the accuracy rate, which is the proportion of correct model predictions among all results whose model predictions are positive; F-Measure represents the harmonic mean of precision (Precision) and recall (Recall); Through the above four indicators, the quantitative results in the confusion matrix can be converted into a ratio between 0 and 1, which is more conducive to standardized measurement.

[0134] Specifically, the performance of the model proposed in the present invention is evaluated, and the performance of the model is analyzed when the seven drilling characteristic parameters such as the inlet and outlet flow difference of the drilling fluid, the inlet and outlet temperature difference, the standpipe pressure, and the total volume of the mud pool are converted into different dimensions. There are a total of 7 time series characteristic parameters, so the maximum dimension of the data is 7 and the minimum data dimension is 2. The influence of data amplification and other sampling methods on the model classification results by analyzing the CycleGAN model is carried out. At the same time, the constructed model is tested multiple times to verify the performance of the model. The results given here are the average values ​​after multiple tests. The classification performance test results of the data set constructed by the present invention when converted to different dimensions are shown in Table 3.

[0135] Table 3 Classification results

[0136]

[0137] Specifically, it can be seen from Table 3 that the model constructed by the present invention has good classification performance, and the classification accuracy of the model increases with the increase of the dimension after feature conversion. When the dimension is 7, the model shows the best classification performance, and the corresponding evaluation indicators are: Accuracy = 0.995, Recall = 0.995, Precision = 0.941, F-Measure = 0.967, which shows that the model has a very high recognition accuracy rate for overflow samples and can provide a strong reference for early overflow monitoring; at the same time, the classification performance of the LCM model of the present invention and other methods are compared, including the introduction of undersampling method (TomekLinks), oversampling method (Borderline-SMOTE) and hybrid sampling method (Tomek Links + Borderline-SMOTE) instead of CycleGAN model into the model constructed by the present invention. The test results of the oversampling method, undersampling method and hybrid sampling method are shown in Tables 4 to 6.

[0138] Table 4 Test results of undersampling method

[0139]

[0140] Table 5 Oversampling method test results

[0141]

[0142] Table 6 Mixed sampling test results

[0143]

[0144]

[0145] Specifically, the classification performance of the best models trained by several algorithms is compared. Figure 6 As shown; wherein, through Table 4-Table 6 and Figure 6 It can be seen that the classification accuracy of the undersampling method (Tomek Links), oversampling method (Borderline-SMOTE) and mixed sampling method (Tomek Links+Borderline-SMOTE) is still high, but overall, the classification performance of these three sampling methods for negative samples is higher than their classification performance for positive samples, and the overall classification effect of the LCM model is significantly better than the classification effect of the three sampling methods. Therefore, when facing the classification problem of small sample imbalanced data, the conventional method of copying minority class samples or deleting majority class samples to balance the positive and negative class sample data sets is used in early overflow monitoring. The model constructed in the above way can effectively solve the problem of imbalanced data sets caused by the serious shortage of positive class samples.

[0146] Example 2

[0147] The present application embodiment provides an early overflow intelligent monitoring system based on a multivariate time series, which is applied to an early overflow intelligent monitoring method based on a multivariate time series in any one of the embodiments 1, see Figure 8 ,include:

[0148] The sample acquisition processing module is used to obtain the time series sample data set of sample drilling, and preprocess and extract features of the time series sample data set to obtain a time series feature sample data set corresponding to the time series sample data set. The time series feature sample data set includes an overflow data set and a non-overflow data set.

[0149] Optionally, the initial model includes a feature extraction module, which includes a data preprocessing unit and a feature extraction unit; wherein the sample acquisition processing module may include:

[0150] The preprocessing submodule is used to perform data cleaning, data noise reduction, missing value processing and normalization processing on the sample data in the time series sample data set in sequence through the data preprocessing unit to obtain the feature data set to be extracted corresponding to the time series sample data set.

[0151] The feature extraction submodule is used to capture the multivariate time series features of different time ranges in the time series sample data set through multiple LSTM networks in the feature extraction unit, and obtain multiple time series feature sets of different time ranges.

[0152] The feature fusion submodule is used to fuse multiple time series feature sets and determine the data set obtained after the feature fusion as the time series feature sample data set.

[0153] The sample amplification module is used to amplify the original overflow data in the overflow data set to a preset amount through a preset initial model to obtain an amplified overflow data set.

[0154] Optionally, the initial model further includes a sample set amplification module, and the sample set amplification module includes a generator G and a generator F; wherein the sample amplification module may include:

[0155] The first data submodule is used to input the original non-overflow data in the non-overflow data set into the generator G for processing to generate amplified data.

[0156] The second data submodule is used to input the amplified data and the original overflow data in the overflow data set into the generator F for processing respectively, to generate overflow identification data corresponding to the amplified data, and to generate overflow data to be tested corresponding to the original overflow data.

[0157] The third data submodule is used to compare the overflow identification data with the overflow data to be tested. If the difference between the overflow identification data and the overflow data to be tested is within a preset difference range, the overflow data to be tested is determined to be amplified overflow data. The amplified overflow data set includes the original overflow data and the amplified overflow data.

[0158] Optionally, the sample set amplification module further includes a discriminator Dx and a discriminator Dy, and the sample amplification module may further include:

[0159] The fourth data submodule is used to calculate the first function loss value of the amplified data generated by the generator G according to the original non-overflow data through the discriminator Dy. If the first function loss value is within the first preset range, the amplified data with the first function loss value within the first preset range is input into the generator F for processing.

[0160] The fifth data submodule is used to generate a second function loss value of the overflow data to be tested based on the original overflow data through the discriminator Dx calculation generator F. If the second function loss value is within the second preset range, the overflow data to be tested with the second function loss value within the second preset range is compared with the overflow identification data.

[0161] Optionally, the sample set amplification module may further include:

[0162] A sixth data submodule, configured to calculate a total function loss value of the sample set amplification module according to the first function loss value and the second function loss value;

[0163] The seventh data submodule is used to amplify the original overflow data in the overflow data set to a preset number through a sample set amplification module that meets the amplification conditions if the total function loss value meets the preset amplification conditions; if the total function loss value does not meet the amplification conditions, adjust the module parameters of the sample set amplification module, and recalculate the total function loss value based on the adjusted module parameters until the total function loss value meets the amplification conditions.

[0164] The model training module is used to input the non-overflow data set and the amplified overflow data set into the initial model, and train the initial model to obtain a prediction model that corresponds to the initial model and is trained.

[0165] Optionally, the initial module further includes a classification module; wherein the model training module may include:

[0166] A data processing submodule, used to input the non-overflow data set and the amplified overflow data set into the classification module for processing, to obtain a first processing result corresponding to the non-overflow data set and a second processing result corresponding to the amplified overflow data set;

[0167] The loss value determination submodule is used to determine whether the initial model meets the preset training end conditions based on the first processing result and the second processing result. If the initial model meets the training end conditions, the initial model that meets the training end conditions is determined as the prediction model. If the initial model does not meet the training end conditions, the model parameters of the initial model are adjusted, and the initial model is continued to be trained based on the adjusted model parameters until the first processing result and the second processing result meet the training end conditions.

[0168] The model prediction processing module is used to obtain the test data set of the well to be tested, and input the test data set into the prediction model for processing to obtain the prediction results corresponding to the well to be tested, and the prediction results include overflow events and non-overflow events.

[0169] Example 3

[0170] The present application embodiment provides an electronic device, such as Fig. 9As shown, it includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, any method in Embodiment 1 is implemented.

[0171] Example 4

[0172] An embodiment of the present application provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable a computer to execute any one of the methods in Example 1.

[0173] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An early overflow intelligent monitoring method based on multivariate time series, characterized in that: The following steps are involved: Acquire a time series sample data set of sample drilling, and perform preprocessing and feature extraction processing on the time series sample data set to obtain a time series feature sample data set corresponding to the time series sample data set, wherein the time series feature sample data set includes an overflow data set and a non-overflow data set; Amplifying the original overflow data in the overflow data set to a preset amount through a preset initial model to obtain the amplified overflow data set; Inputting the non-overflow data set and the amplified overflow data set into the initial model, and training the initial model to obtain a prediction model that corresponds to the initial model and has been trained; A test data set of the well to be tested is obtained, and the test data set is input into the prediction model for processing to obtain a prediction result corresponding to the well to be tested, wherein the prediction result includes overflow events and non-overflow events.

2. The early overflow intelligent monitoring method based on multivariate time series according to claim 1 is characterized in that: The initial model includes a feature extraction module, and the feature extraction module includes a data preprocessing unit and a feature extraction unit; wherein the preprocessing and feature extraction processing are performed on the time series sample data set to obtain a time series feature sample data set corresponding to the time series sample data set, and the time series feature sample data set includes an overflow data set and a non-overflow data set, including: The data preprocessing unit sequentially performs data cleaning, data noise reduction, missing value processing and normalization processing on the sample data in the time series sample data set to obtain a feature data set to be extracted corresponding to the time series sample data set; Through multiple LSTM networks in the feature extraction unit, the multivariate time series features of different time ranges in the time series sample data set are captured respectively, and multiple time series feature sets of different time ranges are obtained; The plurality of time series feature sets are subjected to feature fusion, and a data set obtained after the feature fusion is determined as a time series feature sample data set.

3. The early overflow intelligent monitoring method based on multivariate time series according to claim 2 is characterized in that: The initial model further includes a sample set amplification module, and the sample set amplification module includes a generator G and a generator F; wherein, the original overflow data in the overflow data set is amplified to a preset amount through a preset initial model to obtain the amplified overflow data set, including: Inputting the original non-overflow data in the non-overflow data set into the generator G for processing to generate augmented data; Inputting the amplified data and the original overflow data in the overflow data set into the generator F for processing respectively, generating overflow identification data corresponding to the amplified data, and generating overflow data to be tested corresponding to the original overflow data; The overflow identification data is compared with the overflow data to be tested. If the difference between the overflow identification data and the overflow data to be tested is within a preset difference range, the overflow data to be tested is determined to be amplified overflow data, and the amplified overflow data set includes original overflow data and amplified overflow data.

4. The early overflow intelligent monitoring method based on multivariate time series according to claim 3 is characterized in that: The sample set amplification module further includes a discriminator Dx and a discriminator Dy, and the method further includes: Calculate the loss value of the first function of the generator G to generate the amplified data according to the original non-overflow data through the discriminator Dy, and if the first function loss value is within the first preset range, input the amplified data within the first preset range of the first function loss value into the generator F for processing; The generator F generates a second function loss value of the overflow data to be tested based on the original overflow data through the discriminator Dx. If the second function loss value is within a second preset range, the overflow data to be tested with the second function loss value within the second preset range is compared with the overflow identification data.

5. The early overflow intelligent monitoring method based on multivariate time series according to claim 4 is characterized in that: The method further comprises: Calculate the total function loss value of the sample set amplification module according to the first function loss value and the second function loss value; If the total function loss value meets the preset amplification conditions, the original overflow data in the overflow data set is amplified to a preset quantity through the sample set amplification module that meets the amplification conditions; if the total function loss value does not meet the amplification conditions, the module parameters of the sample set amplification module are adjusted, and the total function loss value is recalculated based on the adjusted module parameters until the total function loss value meets the amplification conditions.

6. The early overflow intelligent monitoring method based on multivariate time series according to claim 5 is characterized in that: The total function loss value is expressed by a first formula, which is: L(G,F,D X ,D Y )=L GAN (G,D Y ,N N ,N P )+L GAN (F,D X ,N P ,N N )+L cGAN (x,y,z)+λL cyc (G,F)+L1(x,y c ); In the formula, L(G,F,D X ,D Y ) represents the total function loss value, L GAN (G,D Y ,N N ,N P ) represents the loss value of the first function, L GAN (F,D X ,N P ,N N ) represents the loss value of the second function, L cGAN (x,y,z) represents the noise loss value, λL cyc (G, F) represents the cycle consistency loss value, L1(x,y c ) represents the difference between the overflow identification data and the overflow data to be tested, and λ represents an adjustable parameter.

7. The early overflow intelligent monitoring method based on multivariate time series according to claim 5 is characterized in that: The initial module also includes a classification module; wherein the non-overflow data set and the amplified overflow data set are input into the initial model, and the initial model is trained to obtain a prediction model corresponding to the initial model and trained, including: Inputting the non-overflow data set and the amplified overflow data set into the classification module for processing respectively, to obtain a first processing result corresponding to the non-overflow data set and a second processing result corresponding to the amplified overflow data set; Whether the initial model satisfies a preset training end condition is determined according to the first processing result and the second processing result; if the initial model satisfies the training end condition, the initial model satisfying the training end condition is determined as a prediction model; if the initial model does not satisfy the training end condition, the model parameters of the initial model are adjusted, and the initial model is continuously trained based on the adjusted model parameters until the first processing result and the second processing result satisfy the training end condition.

8. An early overflow intelligent monitoring system based on multivariate time series, applied to an early overflow intelligent monitoring method based on multivariate time series according to any one of claims 1 to 7, characterized in that: include: A sample acquisition processing module is used to acquire a time series sample data set of sample drilling, and perform preprocessing and feature extraction processing on the time series sample data set to obtain a time series feature sample data set corresponding to the time series sample data set, wherein the time series feature sample data set includes an overflow data set and a non-overflow data set; A sample amplification module, used to amplify the original overflow data in the overflow data set to a preset amount through a preset initial model to obtain the amplified overflow data set; A model training module, used for inputting the non-overflow data set and the amplified overflow data set into the initial model, and training the initial model to obtain a prediction model corresponding to the initial model and trained; The model prediction processing module is used to obtain a test data set of the well to be tested, and input the test data set into the prediction model for processing to obtain a prediction result corresponding to the well to be tested, wherein the prediction result includes overflow events and non-overflow events.

9. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the method according to any one of claims 1 to 7 is implemented when the processor executes the computer program.

10. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, which enable a computer to execute the method of any one of claims 1-7.

Images