Deep and far sea ultra-deep water drilling risk early warning method based on multi-algorithm fusion

Abstract

The invention relates to a deep and far sea ultra-deep water drilling risk early warning method based on multi-algorithm fusion, and the method comprises the steps: S1, obtaining the original time sequence data of a plurality of preset physical parameters of a to-be-early-warned drilling well; s2, performing preprocessing of signal decomposition and denoising processing on the original time sequence data of each physical parameter to obtain corresponding data to be analyzed; s3, for the to-be-analyzed data of each physical parameter, predicting a corresponding risk category and a probability thereof based on a pre-constructed time sequence prediction model; and S4, integrating the risk categories corresponding to all the preset types of physical parameters and the probabilities thereof, and calculating the finally predicted risk category and probability based on an evidence theory.

Description

Technical Field

[0001] The present invention relates to the technical field of deep-sea oil and gas exploitation, and in particular to a deep-sea and ultra-deep-water drilling risk early warning method based on multi-algorithm fusion. Background Art

[0002] With the advancement of technology, oil and gas development has expanded to deep sea areas. Deepwater drilling may lead to some risk accidents, among which leakage and overflow are relatively typical. The occurrence of leakage will cause the suspension or interruption of drilling work, consuming a lot of manpower, material and financial resources. If leakage is not handled in time, it will also cause drilling well collapse, blowout and stuck drill accidents. The occurrence of overflow may also cause blowout, endangering the safety of people's lives and property and damaging the environment.

[0003] The industry requires early, efficient and accurate prediction and warning of well leakage and overflow risk events. At present, downhole overflow and well leakage detection technologies include sensor technology, data processing and analysis, models and algorithms, drilling parameter detection and monitoring, etc. At present, there are still some shortcomings in downhole leakage and overflow warning. The specific situation is as follows: the detection sensitivity is limited, and the current warning technology can only warn when overflow and well leakage occur to a certain extent. The detection sensitivity of minor drilling overflow and well leakage needs to be improved; the warning accuracy needs to be improved: in the drilling process under complex geological conditions, the warning accuracy faces challenges; data processing still needs to be improved: when processing and analyzing a large amount of data, more efficient and accurate algorithms are needed to achieve fast and accurate warnings; limited scope of application: the current warning technology still has some limitations on the scope of application. For different types of wells and different drilling environments, it is necessary to select and adopt different warning technologies in a targeted manner. Summary of the invention

[0004] The present invention provides a deep-sea and ultra-deepwater drilling risk warning method based on multi-algorithm fusion, which can efficiently and accurately predict and warn of the risks of well leakage and overflow in the deepwater drilling process.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, the present application provides a deep-sea and ultra-deepwater drilling risk warning method based on multi-algorithm fusion, the method comprising:

[0007] S1, obtaining original time series data of multiple preset physical parameters of the drilling to be warned;

[0008] S2, preprocessing the original time series data of each physical parameter by signal decomposition and denoising to obtain the corresponding data to be analyzed;

[0009] S3, for each physical parameter to be analyzed, based on the pre-built time series prediction model, predict the corresponding risk category and its probability;

[0010] S4, comprehensively considers the risk categories and probabilities corresponding to all preset types of physical parameters, and calculates the final predicted risk category and probability based on evidence theory.

[0011] In one implementation, in S1, multiple preset physical parameters include: the big hook load of the drilling to be warned, the standpipe pressure, the mud pump pump stroke, the mud pool volume, the mud tank volume, the relative outlet flow rate of the drilling fluid, the outlet temperature, the rotation speed, the outlet drilling fluid density, the inlet drilling fluid density, and the inlet flow rate of the drilling fluid.

[0012] In one implementation, in S2, the preprocessing includes:

[0013] Based on the CEEMDAN algorithm, the original time series data is modally decomposed;

[0014] The signal after modal decomposition is denoised based on wavelet basis functions.

[0015] In one implementation, the specific calculation process of the CEEMDAN algorithm includes:

[0016] (1) The expression of the time domain noise sequence is as follows:

[0017] S i (t) = S (t) + ε0w i (t), i=1,2...K

[0018] (2) Find the maximum and minimum points of the original signal S(t), and then fit these extreme points through the curve interpolation method to obtain the upper envelope S of the signal max (t) and the lower envelope S min (t); calculate the average value of the upper and lower envelopes, and the calculation expression is as follows:

[0019]

[0020] (3) Subtract the original signal S(t) from the average envelope m1(t) to obtain the residual signal d1(t). Reconstruct the time domain signal sequence of the residual signal and calculate the average value of the upper and lower envelopes to SD to obtain the first-order IMF component. Calculate the average value of its signal amplitude to obtain the first IMF component. The calculation expression is as follows:

[0021]

[0022] (5) Calculate the first residual

[0023] (6) Given the residual result, r1(t)+ε1EMD1(w i (t)) reconstruction, where EMD1(w i (t)) is the first modal component of EMD;

[0024] (7) The formula for subtracting the first IMF from the original signal is as follows:

[0025]

[0026] (8) According to the same rule, the average value of the K+1 order residual and the K+1 order IMF is calculated until the residual signal is no longer decomposed:

[0027]

[0028] (9) The final original signal can be expressed as follows:

[0029]

[0030] In one implementation, denoising is performed based on a wavelet basis function, including selection of a wavelet basis function and a threshold function.

[0031] In one implementation, the time series prediction model is a neural network model using a CNN-LSTM network architecture.

[0032] In one implementation, the neural network model is used for the network part of CNN convolution processing, including an input layer, a convolution layer, a pooling layer and a fully connected layer. The input layer inputs the data after noise reduction processing, the convolution layer extracts the input variable features, the pooling layer reduces the data feature dimension, and the fully connected layer connects the current layer with the next layer, as well as the neurons in each section, and calculates the output through the activation function.

[0033] In one implementation, the LSTM network part of the neural network model includes a storage unit and three control gates, namely an input gate, a forget gate, and an output gate; the calculation steps are as follows:

[0034] (1) Forget gate f t The calculation formula is as follows:

[0035]

[0036] Where: W f and b f Respectively represent the forget gate weight matrix and bias; σ is the activation function, and its value range is [0, 1];

[0037] (2) Use the input gate to determine how much new data information to add based on the current input and the state of the memory unit at the previous moment. The tanh activation function generates new information; the calculation formula is as follows:

[0038]

[0039] Where: W i and b i Respectively represent the weight and bias of the input gate; W c and b c are the weight and bias of the memory node respectively; the value range of tanh(x) is (0, 1);

[0040] (3) The memory unit is updated based on the previous state and new information, and the old cell state C t-1 Update to the new cell state C t ; Change the old cell state C t-1 With the attenuation coefficient f t Multiply, delete the data information that is determined to be deleted, and add C t , and get the new cell state C t ; The calculation formula is as follows:

[0041]

[0042] (6) The output gate determines how much data information to output based on the new cell state. t Indicates the degree to which historical information and current input information are initially combined, and outputs the prediction vector h t The calculation formula is as follows:

[0043]

[0044] h t =O t ×tanh(C t )

[0045] Where: W o and b o Represent the weight and bias of the output gate respectively.

[0046] In one implementation, the step S4 includes:

[0047] Define the event set based on the predicted risk category and its probability corresponding to each physical parameter;

[0048] Based on the defined event set, the decision distance of the prediction results of different physical parameters is calculated;

[0049] A similarity matrix is ​​constructed based on the decision distance, and the final fusion probability value is calculated using the support vector, reliability vector, and average trust distribution function.

[0050] The technical solution of the present invention predicts the early leakage risk of drilling through CEEMDAN-wavelet threshold denoising and establishing a CNN-LSTM time series prediction model. According to the change law of the physical parameters of logging when the overflow and well leakage risks occur, the risk probability is preliminarily calculated, and the risk probability results of multiple physical parameters are fused through the DS multi-source parameter fusion method for early warning, so as to prevent drilling blowout accidents and ensure the safe operation of the drilling platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is an overall flow chart of CEEMDAN in an embodiment of the present invention;

[0052] Figure 2 is a flow chart of a data processing method based on CEEMDAN-wavelet threshold denoising according to an embodiment of the present invention;

[0053] Figure 3 LSTM unit structure diagram in an embodiment of the present invention;

[0054] Figure 4 It is a structural diagram of the CNN-LSTM prediction and early warning method of an embodiment of the present invention;

[0055] Figure 5 It is a diagram of the DS multi-source parameter fusion steps of the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0057] In view of the defects and problems of the prior art, the present application provides a deep-sea ultra-deepwater drilling risk warning method based on multi-algorithm fusion, the method comprising:

[0058] S1, obtaining original time series data of multiple preset physical parameters of the drilling to be warned;

[0059] Specifically, in this step, a certain offshore high-temperature and high-pressure deepwater well with a water depth of 87 meters, a core elevation of 23 meters, and a designed well depth of 4063 meters is used as the background of the case study, and the recording method in the work log is to record the real-time data of each characteristic parameter every 5 seconds. The studied characteristics include hook load, standpipe pressure, No. 2 mud pump pump stroke, No. 3 mud pump pump stroke, mud pool volume, mud tank volume, drilling fluid relative outlet flow rate, outlet temperature, and add four control parameters: rotation speed, outlet drilling fluid density, inlet drilling fluid density, and drilling fluid inlet flow rate.

[0060] S2, preprocessing the original time series data of each physical parameter by signal decomposition and denoising to obtain the corresponding data to be analyzed;

[0061] Specifically, this step includes CEEMDAN data decomposition and wavelet threshold denoising. They are described in detail as follows:

[0062] CEEMDAN data breakdown:

[0063] 1000 groups of Gaussian white noise signals with a standard deviation of 0.2 are added to the original signal. The noise standard deviation and implementation control parameters in CEEMDAN are set to 1000 and 0.2 respectively, and the maximum number of screening iterations allowed is set to 200.

[0064] The original signal S(t) is combined with the white noise time domain signal sequence to obtain the noisy time domain signal sequence S i (t), ε0 is the standard deviation of noise that controls the signal-to-noise ratio. The expression of the time domain noise sequence is as follows:

[0065] S i (t) = S (t) + ε0w i (t), i=1,2...K

[0066] Find the maximum and minimum points of the original signal S(t), and then fit these extreme points through the curve interpolation method to obtain the upper envelope S of the signal max (t) and the lower envelope S min (t).

[0067] To find the average value of the upper and lower envelopes, the calculation expression is as follows:

[0068]

[0069] Subtract the original signal S(t) from the average envelope m1(t) to obtain the remaining signal d1(t). Reconstruct the time domain signal sequence for the remaining signal and calculate the average of the upper and lower envelopes until SD is obtained to obtain the first-order IMF component. Calculate the average of its signal amplitude to obtain the first IMF component. The calculation expression is as follows:

[0070]

[0071] Calculate the first residual

[0072] Given the residual result, r1(t)+ε1EMD1(w i (t)) is reconstructed according to the method in step 1, where EMD1(w i (t)) is the first modal component of the EMD.

[0073] The formula for subtracting the first IMF expression from the original signal is as follows:

[0074]

[0075] According to the same rule, the order residual and the average value of the K+1 order IMF are calculated until the residual signal is no longer decomposed. The relevant calculation expression is as follows:

[0076]

[0077]

[0078] The final original signal can be expressed as follows:

[0079]

[0080] Wavelet threshold denoising:

[0081] The wavelet threshold denoising method selects the first few components with relatively large noise in the CEEDMDAN decomposition curve through the autocorrelation method, selects the appropriate threshold for the noisy IMF component, and selects the wavelet basis function db8 and 7 layers by comparison to denoise the data of the characteristic parameters.

[0082] The wavelet coefficients are reconstructed according to the low-frequency decomposition coefficients and the high-frequency threshold coefficients, thereby obtaining a smooth denoised signal.

[0083] S3, for each physical parameter to be analyzed, based on the pre-built time series prediction model, predict the corresponding risk category and its probability;

[0084] Specifically, the time series prediction model is a neural network model using the CNN-LSTM network architecture, including a CNN part and an LSTM part.

[0085] CNN Convolution:

[0086] The well logging data preprocessed by CEEMDAN-wavelet threshold denoising were input into the prediction model.

[0087] Set the step size of the convolution kernel to 1, initialize the CNN weights, extract the features of the input variables through the convolution kernel, and calculate the specific features inside the region.

[0088] The parameters set for the pooling layer are the same as those for the convolutional layer, which reduces the dimension of data features to reduce the number of parameters.

[0089] Each section in the current layer is connected to the neurons through weights, connected to each section in the next layer, and the output is calculated through the activation function.

[0090] LSTM prediction:

[0091] (1) Input the output data into the LSTM model, use the input gate to determine how much new data information to add based on the current input and the memory cell state at the previous moment, and generate new information through the tanh activation function. The memory cell is updated based on the state at the previous moment and the new information, and the output gate determines how much data information to output based on the new cell state. Finally, according to the back-propagation method, the weights and biases are continuously optimized to obtain a stable LSTM neural network model.

[0092] (2) First, the slope of the data curve within 5 minutes is calculated, and the entire data segment is calculated in sequence using a dynamic advancement method. The probability of overflow and well leakage risks is preliminarily characterized based on the slope of the data curve of the characteristic parameters.

[0093] S4, comprehensively considers the risk categories and probabilities corresponding to all preset types of physical parameters, and calculates the final predicted risk category and probability based on evidence theory.

[0094] Specifically, DS (Dempster-Shafer) evidence theory is an uncertainty reasoning method that deals with uncertainty and incomplete information. The core concepts are "evidence" and "combination". In DS theory, "evidence" refers to data containing uncertain information, and "combination" refers to the fusion of this evidence information. It provides a method to fuse information from multiple evidence sources into a whole to improve the accuracy and reliability of the data.

[0095] Step 1: Define the event set as E = [E1, E2, L, E e ] Select P p The probability corresponding to the characteristic parameters that characterize the overflow event, subscript p = 1, 2, L, v, L, w.

[0096] Where E e Representative working conditions include drilling down, pulling out, drilling in, etc.

[0097] Step 2: Calculate the decision distance d between overflow probability results vw , the calculation expression is as follows:

[0098] d vw =2|(P v -0.5)(P v -P w )

[0099] Step 3: Use decision distance d vw Represents the similarity matrix element r vw , construct the similarity matrix R, the expression is as follows:

[0100] r vw =1-d vw (0<r vw <1)

[0101]

[0102] Step 4: Use the support vector H(P p ), reliability vector α p , average trust distribution function value Comprehensive judgment result P(E e ) is used to deduce and calculate a relatively accurate fusion probability value. The calculation formula is as follows:

[0103]

[0104] In the several embodiments provided by the present invention, it should be understood that the disclosed method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0105] The above-mentioned integrated unit implemented in the form of a software functional unit can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor (Processor) to perform some steps of the above-mentioned method of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program codes.

[0106] The above are only preferred embodiments of the present invention and are not intended to limit 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. A deep sea and ultra-deep water drilling risk warning method based on multi-algorithm fusion, characterized in that: The method comprises: S1, obtaining original time series data of multiple preset physical parameters of the drilling to be warned; S2, preprocessing the original time series data of each physical parameter by signal decomposition and denoising to obtain the corresponding data to be analyzed; S3, for each physical parameter to be analyzed, based on the pre-built time series prediction model, predict the corresponding risk category and its probability; S4, comprehensively considers the risk categories and probabilities corresponding to all preset types of physical parameters, and calculates the final predicted risk category and probability based on evidence theory.

2. The deep-sea and ultra-deepwater drilling risk early warning method based on multi-algorithm fusion according to claim 1 is characterized in that: In S1, a variety of preset physical parameters include: the hook load of the drilling to be warned, the standpipe pressure, the mud pump stroke, the mud pool volume, the mud tank volume, the relative outlet flow rate of the drilling fluid, the outlet temperature, the rotation speed, the outlet drilling fluid density, the inlet drilling fluid density, and the inlet flow rate of the drilling fluid.

3. The deep sea and ultra-deep water drilling risk early warning method based on multi-algorithm fusion according to claim 2 is characterized in that: In S2, the preprocessing includes: Based on the CEEMDAN algorithm, the original time series data is modally decomposed; The signal after modal decomposition is denoised based on wavelet basis functions.

4. The deep-sea and ultra-deepwater drilling risk early warning method based on multi-algorithm fusion according to claim 3 is characterized in that: The specific calculation process of the CEEMDAN algorithm includes: (1) The expression of the time domain noise sequence is as follows: S i (t)=S(t)+ε0w i (t),i=1,2…K (2) Find the maximum and minimum points of the original signal S(t), and then fit these extreme points through the curve interpolation method to obtain the upper envelope S of the signal max (t) and the lower envelope S min (t); calculate the average value of the upper and lower envelopes, and the calculation expression is as follows: (3) Subtract the original signal S(t) from the average envelope m1(t) to obtain the remaining signal d1(t). Reconstruct the time domain signal sequence of the remaining signal and calculate the average value of the upper and lower envelopes to SD to obtain the first-order IMF component. Calculate the average value of its signal amplitude to obtain the first IMF component. The calculation expression is as follows: (5) Calculate the first residual (6) Given the residual result, r1(t)+ε1EMD1(w i (t)) reconstruction, where EMD1(w i (t)) is the first modal component of EMD; (7) The formula for subtracting the first IMF from the original signal is as follows: (8) According to the same rule, the average value of the K+1 order residual and the K+1 order IMF is calculated until the residual signal is no longer decomposed: (9) The final original signal can be expressed as follows:

5. The deep sea and ultra-deep water drilling risk early warning method based on multi-algorithm fusion according to claim 3 is characterized in that: De-noising is performed based on wavelet basis functions, including the selection of wavelet basis functions and threshold functions.

6. The deep sea and ultra-deep water drilling risk early warning method based on multi-algorithm fusion according to claim 1 is characterized in that: The time series prediction model is a neural network model using a CNN-LSTM network architecture.

7. The deep sea and ultra-deep water drilling risk early warning method based on multi-algorithm fusion according to claim 6 is characterized in that: The neural network model is used for the network part of CNN convolution processing, including input layer, convolution layer, pooling layer and fully connected layer. The input layer inputs the data after noise reduction processing, the convolution layer extracts the input variable features, the pooling layer reduces the data feature dimension, and the fully connected layer connects the current layer with the next layer, as well as the neurons in each section, and calculates the output through the activation function.

8. The deep sea and ultra-deep water drilling risk early warning method based on multi-algorithm fusion according to claim 6 is characterized in that: The LSTM network part of the neural network model includes a storage unit and three control gates, namely an input gate, a forget gate, and an output gate; the calculation steps are as follows: (1) Forget gate f t The calculation formula is as follows: Where: W f and b f Respectively represent the forget gate weight matrix and bias; σ is the activation function, and its value range is [0,1]; (2) Use the input gate to determine how much new data information to add based on the current input and the state of the memory unit at the previous moment; generate new information using the tanh activation function; the calculation formula is as follows: Where: W i and b i Respectively represent the weight and bias of the input gate; W c and b c are the weight and bias of the memory node respectively; the value range of tanh(x) is (0,1); (3) The memory unit is updated based on the previous state and new information, and the old cell state C t-1 Update to the new cell state C t ; Change the old cell state C t-1 and attenuation coefficient f t Multiply, delete the data information that is determined to be deleted, and add C t , and get the new cell state C t ; The calculation formula is as follows: (6) The output gate determines how much data information to output based on the new cell state. t Indicates the degree to which historical information and current input information are initially combined, and outputs the prediction vector h t , the calculation formula is as follows: h t =O t ×tanh(C t ) Where: W o and b o Represent the weight and bias of the output gate respectively.

9. The deep sea and ultra-deep water drilling risk early warning method based on multi-algorithm fusion according to claim 1 is characterized in that: The S4 includes: Define the event set based on the predicted risk category and its probability corresponding to each physical parameter; Based on the defined event set, the decision distance of the prediction results of different physical parameters is calculated; A similarity matrix is ​​constructed based on the decision distance, and the final fusion probability value is calculated using the support vector, reliability vector, and average trust distribution function.

Images