A method for generating big data of intelligent stratigraphic correlation based on wavelet transform
Through the combination of wavelet transformation and LSTM model, the problems of low prediction accuracy and low efficiency in automatic comparison of well logging data formations in the existing technology are solved, and higher accuracy multi-class data fusion and batch formation division are achieved, which improves the The processing efficiency of well logging data reduces the influence of human subjective factors.
Patent Information
- Application Number
- CN202111302827.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-05
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-11-05
AI Technical Summary
In the automatic comparison of logging data formations in the existing technology, the prediction accuracy is not high, especially at the small-level level, it is difficult to effectively consider the front and back correlation of the logging curve and the fusion of multiple types of data. The traditional method is inefficient and difficult to compare. , affected by human subjective factors and cannot achieve batch rapid stratigraphic division.
Using the intelligent formation comparison big data generation method based on wavelet transform, the logging curve is processed through CoifN wavelet transform, and multi-task training is carried out in combination with the LSTM model, which can better consider the data's front-to-back correlation and inputs of different lengths to build a multi-task structure. Improve the accuracy of the stratification results, and filter the sensitive curves through lithologic intercourse maps to form an efficient data set for training.
It improves the hierarchical prediction accuracy of well logging data, can effectively combine multi-category data analysis, reduces human subjective interference, realizes batch operations, improves the efficiency of strata division, better explores data characteristics, and is faithful to geological information.
Smart Images

Figure CN114021703B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of automatic formation comparison methods for well logging data, and in particular to a method for generating intelligent formation comparison big data based on wavelet transform. Background Art
[0002] The traditional logging data stratigraphic automatic correlation method was first implemented by Rudman and Lankston by matching the vertical layers of two logging signals in a sliding window based on the cross-correlation function. This method is based on the calculation of statistical quantities and can be classified as a mathematical statistics method. In addition, the maximum likelihood estimation method, cluster analysis method, and gray correlation method have also been used in stratigraphic correlation related research. However, mathematical statistics methods have strict requirements on data in practical applications, and their results lack geological information fusion. In current related research, traditional automatic stratigraphic correlation methods can also be divided into non-mathematical statistics methods such as wavelet transform method, expert system method, pattern recognition method, etc.
[0003] In recent years, with the boom of artificial intelligence, intelligent stratigraphic division based on neural networks has begun to attract more attention from researchers. Wu and Nyland established an artificial intelligence stratification system based on linear regression model and heuristic algorithm. Ao Dekui et al. used computers instead of manual labor to realize the recognition of stratigraphic interfaces, which can convert the digital quantity of well logging into symbolic quantity, and then perform rule reasoning on this basis to obtain the stratification information hidden in the well logging curve. In addition, a series of artificial intelligence algorithms such as BP neural network and self-organizing map neural network (SOM) have been used to carry out relevant research on automatic stratification. Generally, the essence of stratigraphic division based on well logging data is to extract, segment and classify the curve according to different features, so it can be regarded as a one-dimensional semantic segmentation task.
[0004] like Figure 1 As shown in the figure, the prior art 1: automatic stratigraphic comparison experiment based on convolutional neural network algorithm[J],
[0005] In the prior art 1, the selected data set is an oil-bearing stratum system, which is further divided into 4 oil layer groups (such as S1), 10 sand layer groups (such as S11), and 30 small layers (such as S111). The SP, ML1, and ML2 curves are selected as the input data of the neural network for the target area, that is, the prior art directly selects the logging data for processing. The final result is obtained according to the different scales of the data set, that is, the single-layer prediction of the oil group-level stratigraphic unit, the single-layer prediction of the sand group-level stratigraphic unit, and the single-layer prediction of the small-layer-level stratigraphic unit.
[0006] The neural network of choice in the prior art is the SegNet network, which does not make direct structural modifications to the network, but only sets specific weight parameters as needed.
[0007] Disadvantages of the existing technology 1: The prediction scale is increasing from the oil layer group to the small layer group. The scale of the small layer group has the smallest amount of data and requires a neural network with higher precision prediction capabilities. The existing technology cannot do this when the prediction target is the smallest level, i.e. the small layer, and the prediction accuracy is not high.
[0008] A large number of well logging is required, which means that the quality and quantity of the original data are required to be high.
[0009] When the input data is a one-dimensional curve, the SegNet network cannot well consider the correlation between the data before and after. For the logging curve of each small layer, its characteristics will be affected by the upper and lower geological environment.
[0010] The convolution kernel used is an N*1 array, the number and size of which are determined based on experience. The article does not indicate whether a one-dimensional convolution method is used, that is, the convolution direction of the technology is unclear. Using this convolution kernel for data convolution operations cannot effectively preserve the longitudinal characteristics of the logging data, nor can it better combine the characteristics of the three input logging curves in the same layer (lateral characteristics).
[0011] Prior Art 2: Multi-label Multi-task Learning Based on RNN [D]. China University of Petroleum (Beijing), 2017. The neural network structure here is different from that of the present invention. Figures 2 to 4 shown.
[0012] There are usually two problems in traditional methods: the stratification results are usually affected by the actual experience of experts and are highly subjective;
[0013] Secondly, automatic stratification based on these algorithms often finds it difficult to achieve accurate stratification based on multi-class data fusion.
[0014] In addition, the common disadvantage of traditional methods is that they cannot achieve rapid batch stratum division, and the process is time-consuming and labor-intensive.
[0015] Under complex geological conditions, traditional methods are inefficient and difficult to compare.
[0016] In an existing study, the above-mentioned prior art 1 has problems: lack of data set or low data quality, and high-precision small layer division problem. Summary of the invention
[0017] In order to solve the problems existing in the prior art, the present invention provides an intelligent formation comparison big data generation method based on wavelet transform. The present invention can better consider the front-end correlation of logging data, and it also has the ability to process input data of different lengths. It has a more flexible data processing mechanism, can effectively combine multiple types of data for analysis, can better mine data features, be faithful to data, avoid interference from human subjective factors, can perform batch operations, and improve efficiency.
[0018] The present invention provides a technical solution, a method for generating big data of intelligent stratigraphic comparison based on wavelet transform, comprising the following steps:
[0019] Step 1: Based on the existing data of the target work area, the logging curves sensitive to lithology are screened out by using the lithology cross-plot analysis with the help of the existing logging layer division results;
[0020] Step 2: Select well logging curves from the initial data set;
[0021] Step 3: Perform CoifN wavelet transform on the logging curve;
[0022] Step 4: Determine the size of the regularity after CoifN wavelet transform;
[0023] Step 5: If the regularity is good, the support length will be longer and the calculation time will be longer. At this time, the number of decomposition levels can be determined;
[0024] Step 6: According to the decomposition results of the previous step, combined with the curve characteristics of each small layer, 1-3 high-frequency data are selected and added to the original data set in the first step to form the final training data set;
[0025] Step 7: Build a multi-task LSTM model, and use the data in step 6 as the training data set to train the model and obtain one-dimensional well logging curve data;
[0026] Step 8: Get the network based on the curve data in step 7, and then get the small layer division result based on the network, and use the accuracy formula to evaluate its quality.
[0027] Preferably, the screening condition of step 1 is that two or more different lithologies can be linearly distinguished well in the cross-plot.
[0028] Preferably, the CoifN wavelet transform in step three is a wavelet basis function, and the value range of N is 1-5. The larger N is, the better the denoising effect and smoothness of the processed signal are, and the regularity increases with the increase of N.
[0029] Preferably, the decomposition level in step 5 is 1-8 levels.
[0030] The beneficial effects of the intelligent stratigraphic comparison big data generation method based on wavelet transform of the present invention are as follows:
[0031] 1. As the depth increases, the logging curve has different curve characteristics according to different stratum deposition conditions. The deposition of the stratum is affected by the deposition environment, so the lithology of the stratum is correlated in the deposition sequence. Compared with the SegNet network, the LSTM network structure can better consider the correlation between the logging data and the past and future. It also has the ability to process input data of different lengths and has a more flexible data processing mechanism. Segnet is a fixed input data size.
[0032] 2. The present invention adopts a multi-task structure to obtain more accurate stratification results and can effectively combine multiple types of data for analysis
[0033] 3. The deep learning method is a data-driven method that can better mine data features, be faithful to the data, avoid interference from human subjective factors, and can be operated in batches, thereby improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 The present invention is a flow chart of a method of prior art 1.
[0035] Figure 2 Detailed structure diagram of the multi-task LSTM network.
[0036] Figure 3(a) is a diagram of the first layer structure of a traditional RNN;
[0037] Figure 3(b) is the second layer structure diagram of the traditional RNN.
[0038] Figure 4 This is the LSTM structure diagram.
[0039] Figure 5 This is a lithology cross-plot analysis diagram.
[0040] Figure 6 This is the Coif5 wavelet transform result of the logging curve.
[0041] Figure 7 This is the final training dataset diagram.
[0042] Figure 8 This is a graph of multi-well batch segmentation results (the black line is the expert segmentation result, and the red line is the neural network segmentation result).
[0043] Fig. 9 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0044] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0045] The present invention provides a method for generating big data of intelligent stratigraphic comparison based on wavelet transform.
[0046] 1. First, based on the existing data of the target work area, the existing logging layer division results are used to analyze the lithology cross-plot ( Figure 5 ) Filter out logging curves that are sensitive to lithology, such as wave impedance curve, natural gamma curve, natural potential curve, and microelectrode curve as the initial data set of the neural network (the categories of sensitive logging curves can be 2-3 types). The screening condition is that two or more different lithologies can be linearly distinguished well in the intersection diagram.
[0047] 2. Select logging curves from the initial data set (it can be any type, but for the sake of computational efficiency, there should not be too many), and perform CoifN wavelet transform on the logging curves (Coif is a wavelet basis function, N: 1-5, the larger the N, the better the denoising effect and smoothness of the processed signal (regularity increases with N). The better the regularity, the longer the support length and the longer the calculation time). At this time, it is necessary to determine the level of decomposition (level 1-8), such as: 7-level decomposition, you will get 7 pairs of low-frequency and high-frequency data (a is low frequency and d is high frequency), such as Figure 5 The wavelet basis function used is coif5,
[0048] 3. According to the decomposition results of the previous step, combined with the curve characteristics of each small layer, 1-3 high-frequency data are selected and added to the original data set in the first step to form the final training data set (the wavelet transform process is equivalent to data augmentation), such as Figure 6 As shown,
[0049] 4. Build a multi-task LSTM model and use the data in (3) as the training data set for model training. For one-dimensional well logging curve data, the boundary division of each stratum requires high-precision prediction results to obtain well logging interpretation results that conform to the actual geological conditions.
[0050] The multi-task in this application can be divided into two points: (1) completing the classification task to achieve different layer recognition and (2) completing the regression task to achieve the precise positioning of the specific layer interface. For these two tasks, Figure 1 Some structures in need to be modified; for these two tasks, the neural network model Figure 1 Based on the modification, Figure 1 In the figure, x1…T represents the input data. The logging curve of each well (including data augmentation data) is put into the improved multi-task LSTM model for model training. A curve can be determined by a variety of segmentation methods, such as equal division, segmentation according to formation, segmentation according to sedimentary characteristics, and small layers divided according to needs. The segmented curve becomes x1…T. Since this network is supervised learning, that is, the neural network input and output are known, the input represents the input data in the training data set ( Figure 7 ), the output represents the sub-layer division result (label), the label of the training data set is the sub-layer division result explained by the expert, and when each input curve of the same well is put into the model, their corresponding labels are the same label (because there is only one optimal solution for the sub-layer division of a well).
[0051] Figure 2 In and is the output data of the hidden layer t of LSTM, m represents the regression task, n represents the classification task. In the model structure of task m, the softmax layer is removed, and the loss function for calculating the output data is the MSE loss function, whose formula is L (m) =||y (m) -s (m) || 2 s (m) is the label of the regression task, y (m) The network is the output data; in the n task, the softmax layer is removed, and the loss function for calculating the output data is the cross entropy loss function, whose formula is L (n) = -logSoftmax(y (n) ,S (n) ), s (n) is the label of the classification task, y (n) The network is the output data, and the final joint loss function
[0052]
[0053] Among them, α1 and α2 are parameters used to measure the importance of classification and regression tasks. If α_1 is increased, the proportion of L(m) in the joint loss function will decrease; similarly, if α_1 has a small value, the neural network model will pay more attention to the learning and expression of task m. m is the loss function of the regression task above, L n is the loss function of the above classification task, m is the regression task, and n is the classification task.
[0054] 5. Use the model obtained in 4 to batch divide the wells in the target work area that have not been divided into formations, such as Figure 8 shown.
Claims
1. A method for generating big data for intelligent stratigraphic comparison based on wavelet transform, characterized in that: The following steps are involved: Step 1: Based on the existing data of the target work area, the logging curves sensitive to lithology are screened out by using the lithology cross-plot analysis with the help of the existing logging layer division results; Step 2: Select well logging curves from the initial data set; Step 3: Perform CoifN wavelet transform on the logging curve; Step 4: Determine the size of the regularity after CoifN wavelet transform; Step 5: If the regularity is good, the support length will be longer and the calculation time will be longer. At this time, the number of decomposition levels can be determined; Step 6: According to the decomposition results of the previous step, combined with the curve characteristics of each small layer, 1-3 high-frequency data are selected and added to the original data set in the first step to form the final training data set; Step 7: Build a multi-task LSTM model, and use the data in step 6 as the training data set to train the model and obtain one-dimensional well logging curve data; Step 8: Get the network based on the curve data in step 7, and then get the small layer division result based on the network, and use the accuracy formula to evaluate its quality. The specific formula is: Among them, N pred is the number of correctly classified sample points of the predicted well, N well is the total number of sampling points of the prediction well.
2. The method for generating big data for intelligent stratigraphic comparison based on wavelet transform according to claim 1 is characterized in that: The screening condition of step 1 is that two or more different lithologies can be linearly distinguished well in the cross-plot.
3. The method for generating big data for intelligent stratigraphic comparison based on wavelet transform according to claim 1 is characterized in that: The CoifN wavelet transform in step three is a wavelet basis function, and the value range of N is 1-5. The larger N is, the better the denoising effect and smoothness of the processed signal are, and the regularity increases with the increase of N.
4. The method for generating big data for intelligent stratigraphic comparison based on wavelet transform according to claim 1, characterized in that: The decomposition level in step 5 is 1-8.
