Overflow early warning method of depth feature crossover sequence analysis network
Through the deep feature cross-sequence analysis network, combined with factor decomposition machine and LSTM neural network, the two-dimensional and three-dimensional data characteristics in deep water drilling environment are extracted and fused, and the uncertainty of traditional overflow early warning methods and the lack of early warning capabilities are solved. , achieving more accurate and reliable overflow warning.
Patent Information
- Application Number
- CN202411283210.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-13
- Publication Date
- 2025-01-03
AI Technical Summary
In deep-water drilling environments, traditional overflow warning methods rely on personal experience and have great uncertainty. The existing models lack early warning capabilities and are difficult to accurately predict overflow occurrence in deep-water environments, affecting drilling safety and efficiency.
Deep feature cross-sequence analysis network, combined with factor decomposition machine and long-term memory neural network, extract cross-sequence characteristics and sequence pattern characteristics of two-dimensional and three-dimensional input data, and fuse information through the feature cross-sequence module and sequence analysis module to perform overflow warning classification .
It improves the accuracy and reliability of overflow warning, realizes early warning, improves the timeliness and accuracy of deep-water drilling overflow warning, and enhances the ability to prevent overflow.
Smart Images
Figure CN119249198A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of deepwater oil and gas resource development, and in particular to an overflow early warning method of a deep feature cross-sequence analysis network. Background Art
[0002] In the marine environment of deepwater drilling, multiple factors such as wind, waves, and currents work together to cause the drilling platform to rise and fall and swing, which in turn affects the outlet flow of the drilling fluid and the volume fluctuation of the mud pool. These factors directly affect the accuracy of the mud pool liquid level detection device, thereby increasing the difficulty of early warning of overflows. In addition to sea conditions, low temperature, high pressure, and lack of seabed geological and lithological information in deepwater environments also bring more challenges to overflow warning.
[0003] In terms of overflow warning methods, traditional overflow warnings rely entirely on personal experience and judgment, which has great uncertainty.
[0004] With the rapid development of electronic technology, sensor technology, and artificial intelligence technology, more feasible solutions have been brought to overflow monitoring. For example, some expert systems and Bayesian estimation methods based on mathematical models have been put into use and have achieved certain results.
[0005] However, deepwater well construction is technically difficult, with high investment and operational risks. Overflow prediction is time-sensitive and requires a certain amount of advance warning time to effectively prevent and deal with overflows. Some existing models do not have good early warning capabilities and there is much room for improvement. Summary of the invention
[0006] In view of the above problems, the purpose of the present invention is to provide an overflow warning method based on a deep feature cross-sequence analysis network, which can solve the problems of the prior art, improve the accuracy and reliability of deepwater drilling overflow warning, and achieve early warning.
[0007] To achieve the above object, the present invention adopts the following technical solutions:
[0008] In a first aspect, the present application provides an overflow warning method for a deep feature cross-sequence analysis network, the method comprising:
[0009] Acquire input data of the deepwater drilling to be warned, the input data including two-dimensional input data and three-dimensional input data, the two-dimensional input data is obtained by preprocessing according to the number of parameters and the data volume in the original overflow data, and the three-dimensional input data is obtained by preprocessing according to the number of parameters, the data volume and the time step in the original overflow data;
[0010] Inputting the input data into a pre-constructed deep feature cross-sequence analysis network, wherein the deep feature cross-sequence analysis network includes a feature cross-module and a sequence analysis module;
[0011] The feature intersection module extracts a first intersection feature and a second network feature of the parameters of the two-dimensional input data, and outputs a result of fusing the first intersection feature and the second network feature;
[0012] The sequence analysis module extracts local pattern features of the sequence as output according to the three-dimensional input data;
[0013] The labels of the warning results are classified according to the output of the feature cross-module and the output of the sequence analysis module.
[0014] In one implementation, the feature intersection module includes a factorization machine for extracting intersection features.
[0015] In one implementation, the factor decomposition machine includes an FM layer for extracting low-order features of input data, and a NN layer for extracting high-order features of input data.
[0016] In one implementation, the process of solving the FM layer includes:
[0017]
[0018] Among them, w 0 is the bias term, x i is the value of feature i of the input data, x j is the value of feature j of the input data, w i is the weight of feature i, v i is x i The latent factor vector of the feature, v j is x j The latent factor vector of the feature, <v i ,v j > represents the feature vector v i and the eigenvector v j The inner product of .
[0019] In one implementation, the feature crossover module, the output of the FM layer is input to the NN layer, the purpose is to learn higher-order feature crossover based on the latent vector obtained by FM, and the final output is y FM-NN .
[0020] In one implementation, the feature crossover module further includes an artificial neural network ANN for extracting network features.
[0021] In one implementation, the solution process of the artificial neural network ANN includes:
[0022] The input of ANN is as follows:
[0023] a (0) =[e 1 ,e 2 ,…,e i ,…,e m ]
[0024] Among them, e i represents the embedding of the i-th feature, m represents the total number of features, and a (0) Represents the initialization of the neural network;
[0025] After input to the hidden layer:
[0026] a l+1 =σ(W l a l +b l )
[0027] Among them, l represents the depth of the network layer, σ represents the activation function, and a l represents the output of layer l, W l represents the weight of layer l, b l represents the bias term of layer l;
[0028] The output of ANN is as follows:
[0029] y ANN =W |H|+1 ·a |H| +b |H|+1
[0030] Among them, y DNN represents the output of the network, and |H| represents the number of hidden layers.
[0031] In one implementation, the output of the feature cross-module is expressed as:
[0032] y DFC =concatenate(y FM-N ,y ANN )
[0033] Among them, y DFC It represents the output of the feature cross module, and concatenate represents the feature joint function.
[0034] In one implementation, the sequence analysis module includes an LSTM neural network.
[0035] The present invention adopts the above technical solution, which has the following advantages:
[0036] This method proposes a deep feature cross-sequence analysis network to improve the accuracy and foresight of overflow warning. At the same time, the fused time series model includes a one-dimensional convolutional neural network and a long short-term memory neural network, which can improve the accuracy and reliability of deepwater drilling overflow warning and achieve early warning. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 is a schematic diagram of a network framework of a deep feature cross-sequence analysis network according to an embodiment of the present invention;
[0038] Figure 2 is a schematic diagram of a characteristic cross-connect module of an embodiment of the present invention;
[0039] Figure 3 is a schematic diagram of a sequence analysis module according to an embodiment of the present invention;
[0040] Figure 4 is a graph of experimental results of various models in a calculation example of the present invention;
[0041] Figure 5 This is a comparison chart of the early warning effects of a calculation example. DETAILED DESCRIPTION
[0042] 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.
[0043] In view of the above problems in the prior art, an embodiment of the present invention provides an overflow warning method for a deep feature cross-sequence analysis network, the method comprising:
[0044] S1, obtaining input data of the deepwater drilling to be warned, wherein the input data includes two-dimensional input data and three-dimensional input data;
[0045] Specifically, the two-dimensional input data is obtained by preprocessing according to the number of parameters and the data volume in the original overflow data, and the two-dimensional input data is obtained by preprocessing according to the number of parameters, the data volume and the time step in the original overflow data.
[0046] For example, raw overflow data can be collected by various sensor devices, where the collected parameters include temperature, pressure, flow information, and may also include geographical and geological information. Two-dimensional data is only related to each parameter and its data volume. Three-dimensional data is also related to time series. The raw overflow data obtained by the sensor can be preprocessed to obtain two-dimensional input data and three-dimensional input data.
[0047] S2, inputting the input data into a pre-built deep feature cross-sequence analysis network, wherein the deep feature cross-sequence analysis network includes a feature cross-module and a sequence analysis module;
[0048] S3, the feature cross module extracts cross features and network features of the two-dimensional input data, and outputs the result of fusing the cross features and the network features;
[0049] S4, the sequence analysis module extracts local pattern features of the sequence as output according to the three-dimensional input data;
[0050] S5, classifying the labels of the warning results according to the output of the feature cross-section module and the output of the sequence analysis module.
[0051] The above method is described in a more detailed embodiment below in conjunction with more drawings of the present application.
[0052] In this embodiment, a deep feature cross sequence analysis network (DFCSAN) is pre-constructed. Figure 1 Indicates.
[0053] DFCSAN is mainly composed of two parts: feature cross-link module and sequence analysis module. According to the characteristics of the two modules, the input data of the feature cross-link module is two-dimensional data, and the input data of the sequence analysis module is three-dimensional data. The two modules will be described in detail below.
[0054] The feature cross-module is mainly composed of two components: Factorization Machine (FM) and artificial neural network. FM is responsible for extracting low-order features, and the neural network is responsible for extracting high-order features. The structure of this module is as follows: Figure 2 shown.
[0055] from Figure 2 It can be seen that the input of this module is shared and is two-dimensional data. The right side of the module is the DFM part, which is divided into the FM layer and the NN layer. The calculation formula of the FM layer is shown in formula (1).
[0056]
[0057] Among them, w 0 is the bias term, x i is the value of feature i of a sample, x j is the value of feature j of a sample, w i is the weight of feature i, v i is x i The latent factor vector of the feature, v j is x j The latent factor vector of the feature, <v i ,v j > represents the feature vector v i and the eigenvector v j The inner product of .
[0058] It can be seen that DFM is divided into two parts. The linear part is to assign a weight to each feature and perform weighted summation; the cross part is to multiply the features two by two, then assign weights and perform weighted summation, and then add the two results together as the output of the FM layer. On this basis, the output of the FM layer is input to the NN layer, the purpose is to learn higher-order feature crosses based on the latent vector obtained by FM, and the final output is y FM-NN .
[0059] Figure 2 The right side of is the artificial neural network part in the feature cross module, and the input of this module is shown in formula (2).
[0060] a (0) =[e 1 ,e 2 ,…,e i ,…,e m ] (2)
[0061] Among them, e i represents the embedding of the i-th feature, m represents the total number of features, and a (0) Represents the initialization of the neural network.
[0062] After inputting into the hidden layer, it is shown as formula (3).
[0063] a l+1 =σ(W l a l +b l ) (3)
[0064] Among them, l represents the depth of the network layer, σ represents the activation function, and a l represents the output of layer l, W l represents the weight of layer l, b l Represents the bias term of layer l.
[0065] The final output of the artificial neural network is shown in formula (4).
[0066] y ANN =W |H|+1 ·a |H| +b |H|+1 (4)
[0067] Among them, y DNN represents the output of the network, and |H| represents the number of hidden layers.
[0068] Thus, the output y of the FM part is obtained FM＝NN and the output y of the artificial neural network part ANN , the outputs of the two parts are fused to obtain the output of the feature cross module, as shown in formula (5).
[0069] y DFC =concatenate(y FM-NN ,y ANN ) (5)
[0070] Among them, y DFC It represents the output of the feature cross module, and concatenate represents the feature joint function.
[0071] The obtained feature cross-module output is used by the framework later.
[0072] The best performing LSTM model in the sequence analysis module is composed of time series. The function of LSTM is to memorize and extract local patterns in the sequence, and to capture the long-term dependencies of the data. The structure of this module is as follows Figure 3 shown.
[0073] from Figure 3 As you can see, the input of this module is three-dimensional data in the format of (time step, data volume, number of features). For the LSTM part, the data first enters the LSTM layer to obtain the hidden state H of each time step. T , as shown in formula (6):
[0074] H T =LSTM(X T ) (6)
[0075] Among them, X T is the input data, and LSTM is the LSTM layer. After obtaining the hidden state, it is input into the fully connected layer for analysis, and the output of the LSTM module is obtained, which is also the output of the sequence analysis module, as shown in formula (7):
[0076] y SA =y LSTM =Dense(H T) (7)
[0077] Among them, y LSTM is the output of the LSTM module, y SA It is the output of the sequence analysis module.
[0078] The above describes the data transmission process of the feature cross-module and the sequence analysis module, and obtains the output y of the two modules. DFC and SA The final step of DFCSAN is to combine the outputs of the two modules and identify the overflow through the fully connected layer and the Sigmoid layer. Specifically, as shown in formula (8):
[0079] y Prb-Overflo =concatenate(Dense(y DFC ,y SA )) (8)
[0080] Among them, y Prb-Overflow It represents the probability of overflow occurring.
[0081] Finally, the proposed overflow warning method is used to train and test the overflow warning task, and the performance is compared with that of the existing methods. The results are shown in Table 1 and Figure 4 , 5 shown.
[0082]
[0083] Table 1
[0084] In summary, an overflow warning method based on a deep feature cross-sequence analysis network proposed by model fusion was compared with existing methods. It achieved the best values in terms of recall rate and F1_Score, and was able to achieve early warning of overflow, verifying that the model is capable of performing overflow warning tasks well.
[0085] 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.
[0086] 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.
[0087] 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. An overflow warning method based on a deep feature cross-sequence analysis network, characterized in that: The method comprises: Acquire input data of the deepwater drilling to be warned, the input data including two-dimensional input data and three-dimensional input data, the two-dimensional input data is obtained by preprocessing according to the number of parameters and the data volume in the original overflow data, and the three-dimensional input data is obtained by preprocessing according to the number of parameters, the data volume and the time step in the original overflow data; Inputting the input data into a pre-constructed deep feature cross-sequence analysis network, wherein the deep feature cross-sequence analysis network includes a feature cross-module and a sequence analysis module; The feature intersection module extracts the cross features and network features of the parameters of the two-dimensional input data, and outputs the result of fusing the cross features and the network features; The sequence analysis module extracts local pattern features of the sequence as output according to the three-dimensional input data; The labels of the warning results are classified according to the output of the feature cross-module and the output of the sequence analysis module.
2. The overflow warning method of the deep feature cross-sequence analysis network according to claim 1 is characterized in that: The feature crossover module includes a factor decomposition machine for extracting crossover features.
3. The overflow warning method of the deep feature cross-sequence analysis network according to claim 2 is characterized in that: The factor decomposition machine includes an FM layer for extracting low-order features of input data and a NN layer for extracting high-order features of input data.
4. The overflow warning method of the deep feature cross-sequence analysis network according to claim 3 is characterized in that: The solution process of the FM layer includes: Among them, w0 is the bias term, x i is the value of feature i of the input data, x j is the value of feature j of the input data, w i is the weight of feature i, v i is x i The latent factor vector of the feature, v j is x j The latent factor vector of the feature, <v i ,v j > represents the feature vector v i and the eigenvector v j The inner product of .
5. The overflow warning method of the deep feature cross-sequence analysis network according to claim 4 is characterized in that: The feature cross module, the output of the FM layer is input to the NN layer, the purpose is to learn higher-order feature cross based on the latent vector obtained by FM, and the final output is y FM-NN .
6. The overflow warning method of the deep feature cross-sequence analysis network according to claim 5 is characterized in that: The feature crossover module also includes an artificial neural network ANN for extracting network features.
7. The overflow warning method of the deep feature cross-sequence analysis network according to claim 6 is characterized in that: The solution process of the artificial neural network ANN includes: The input of ANN is as follows: a (0) =[e1,e2,…,e i ,…,the m ] Among them, e i represents the embedding of the i-th feature, m represents the total number of features, and a (0) Represents the initialization of the neural network; After input to the hidden layer: a l+1 =σ(W l a l +b l ) Among them, l represents the depth of the network layer, σ represents the activation function, and a l represents the output of layer l, W l represents the weight of layer l, b l represents the bias term of layer l; The output of ANN is as follows: y ANN =W |H|+1 ·a |H| +b |H|+1 Among them, y DNN represents the output of the network, and |H| represents the number of hidden layers.
8. The overflow warning method of the deep feature cross-sequence analysis network according to claim 7 is characterized in that: The output of the feature cross module is expressed as: and DFC =concatenate(and FM-NN ,and ANN ) Among them, y DFC It represents the output of the feature cross module, and concatenate represents the feature joint function.
9. The overflow warning method of the deep feature cross-sequence analysis network according to claim 1 is characterized in that: The sequence analysis module includes an LSTM neural network.