Stratigraphic division model visualization method and system using confidence score, and terminal thereof

Abstract

The invention discloses a stratigraphic division model visualization method using confidence scores, which comprises the following steps: drawing spatial position coordinates of oil mines and screening similar oil mines to form a data set; performing abnormal value elimination and standardization processing on a logging curve in the data set, generating a two-dimensional representation fragment by using a sliding window, performing vectorization processing on a stratum label, constructing a convolutional neural network model, performing model training by using training set data, and optimizing network parameters through multiple iterations, so that the network can accurately predict the stratum label. And after training is completed, logging data fragments to be visualized are input into the trained network, and a feature map is output in a specified convolutional layer. Carrying out up-sampling on the feature map, carrying out point-by-point multiplication on the feature map and the input fragment, and inputting into the convolutional neural network with the same structure again to obtain a confidence score of the feature map; and taking the confidence score of the feature map of each channel as a weight, and performing linear weighting on the weight and the feature map of the corresponding channel to obtain a class discrimination saliency map.

Description

Technical Field

[0001] The present invention relates to the technical field of oil logging interpretation, and specifically to a visualization method, system and terminal of a formation division model using confidence scores. Background Art

[0002] Scarce oil resources are strategic resources indispensable for a country's survival and development. However, with the emergence of various high-tech technologies, the demand for oil in each country is increasing continuously. In the long run, the world's oil reserves will become less and less, and even dry up. If new oil fields are developed again, the development difficulty is great, the cost is high, and the risk is high. While if the existing oil fields are developed further, the cost and risk are relatively low. This is a practical method to promote the long-term high and stable production of oil fields.

[0003] Well logging interpretation is the process of using various well logging data to identify the reservoirs in the wellbore profile, classify their lithology and fluid types, and determine their geological parameters. Among them, formation division is the most basic and important task. The accurate division of formations is of great significance for subsequent well logging interpretation tasks such as reservoir identification, fluid identification, and lithology division. When surveying existing oil fields, well logging sensors with specific functions will transmit corresponding well logging curve data, which are the most original data in well logging data, record the physical parameters of rock formations changing with downhole depth, and are a direct reflection of geological characteristics. Therefore, directly using well logging curve data to carry out formation division is of great significance for the further development of existing oil fields. At present, formation division mainly relies on the subjective experience of well logging experts for judgment. Due to the excessive subjective factors of the manual interpretation method, there are great differences in formation division conclusions. While if a dedicated deep neural network is constructed for formation division, the errors caused by a large number of human factors such as insufficient subjective experience can be reduced.

[0004] Since the development of formation division to date, the manual interpretation method has formed a complete process. Well logging experts are used to comprehensively analyzing the morphology of well logging curves using personal experience to achieve formation division. Without understanding the specific principle of the application method, most experts are skeptical about intelligent formation division, resulting in the slow development of intelligent formation division. Therefore, it is very important to conduct interpretability research on the application process of deep learning, which can establish trust between well logging experts and deep learning methods so as to better apply them to formation division. At present, the deep learning interpretability for formation division is still in a blank stage. Therefore, conducting visualization research on the formation division task based on deep learning and realizing the mapping of well logging curve features to the feature space can help well logging interpreters better understand its application principle, and then accelerate the application of deep learning algorithms to the formation division task. Summary of the Invention

[0005] In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to provide a method, system and terminal for visualizing a stratigraphic division model using confidence scores, so as to solve the technical problems that the main manual interpretation method for stratigraphic division in the prior art has too many subjective factors and a large number of errors caused by human factors.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for visualizing a stratigraphic partitioning model using confidence scores comprises the following steps:

[0008] Step 1, draw the spatial position coordinates of all oil wells in the block, and select the oil wells with similar spatial position coordinates as a data set, the data set includes several logging curve data;

[0009] Step 2: remove outliers in the well logging curve data in the data set, standardize the well logging curve data of each oil well according to the curve, select a sliding window to obtain a two-dimensional representation fragment of the two-dimensional well logging data, convert all the well logging curve data in the form of discrete data in each well into two-dimensional fragments of the sliding window size, obtain the formation label corresponding to each fragment, vectorize all geological layer labels to obtain a data set, and divide the obtained data set into a training set and a test set according to the ratio;

[0010] Step 3: Construct a convolutional neural network model, input the data in the training set into the convolutional neural network to obtain the network output, compare the network output with the formation label, and calculate the loss function. Then, calculate the gradient of the network parameters to the loss function through the back propagation algorithm, and repeat the training process until the preset maximum number of training rounds is reached to complete the training.

[0011] Step 4: Input the two-dimensional representation fragment of the well logging data to be visualized into the trained convolutional neural network, select the convolutional layer to be visualized, and output the feature map of the layer; upsample the output feature map to the input fragment size and multiply it point by point with the input two-dimensional representation fragment of the well logging data, and then re-input it into the convolutional neural network with the same structure to obtain the confidence score of each feature map;

[0012] Step 5: Use the confidence score of each channel feature map as a weight and linearly weight it with the feature map of the corresponding channel to obtain a class discrimination saliency map;

[0013] Step 6, repeating steps 1 to 5, visualizing multiple two-dimensional characterization segments of logging data of the same formation, obtaining corresponding class discrimination saliency maps and obtaining mapping results of the electrical characteristics of the formation in the feature space.

[0014] Preferably, in step 1, the hierarchical clustering method is used to preferentially screen oil wells with similar spatial coordinates as the data set, and the logging curve data in the data set are correspondingly used as the data sources of several features. The logging curve data are presented in the form of discrete data at a single depth point, and each depth point corresponds to a formation attribute.

[0015] Preferably, several logging curve data include Depth data, natural gamma ray (GR) data, spontaneous potential (SP) data, acoustic travel time (AC) data, array induction resistivity (AT90) data, and array induction resistivity (AT20) data.

[0016] Preferably, in step 2, the logging curve data of each oil well are standardized according to the curve. The standardization is Z-Score standardization, and the formula is as follows:

[0017]

[0018] where μ is the data mean and σ is the data standard deviation.

[0019] Preferably, in step 2, the size of the sliding window is selected as 96×6 pixels and slides along the direction of the logging curve data at an interval of 1 depth point. Each time, a two-dimensional characterization segment of the logging data with a size of 96×6 can be obtained within the window coverage; the logging curve data in the form of discrete data in each well are all converted into two-dimensional segments with a size of 96×6. The formation label corresponding to each segment is the formation label that appears most frequently among the 96 depth points selected by the window. Finally, all the geological layer labels are vectorized. The specific process of vectorizing all the geological layer labels is as follows:

[0020] Count the types n of formation labels in the training set, number the formation labels from 0, and then vectorize the numbers of each label into n-dimensional vectors, where one index in the vector is 1 and the rest of the indices are 0.

[0021] Preferably, in step 3, when constructing the convolutional neural network model, first design the architecture of the convolutional neural network. The architecture of the convolutional neural network includes 3 convolutional layers, 2 fully connected layers, and a softmax classifier. The number of convolutional kernels in the 3 convolutional layers are 64, 128, and 256 respectively, the size of the convolutional kernel is 3×3, the stride is 1, and the padding method is same; the same padding method is to pad a circle of 0s around the array to increase the size of the array; the number of nodes in both fully connected layers is 512; the number of nodes in the softmax classifier is determined by the counted types of formation labels.

[0022] Preferably, in step 4, the formula for solving the confidence score of the two-dimensional characterization segment to be visualized is as follows:

[0023]

[0024] in, is the output feature map of the kth channel of the lth convolutional layer, for A global confidence score for the 2D representation segment score;

[0025] The calculation formula is:

[0026] Among them, the Up function represents the feature map The upsampled dimension is the same as the input vector, and the Norm function represents the normalization of the upsampled feature map.

[0027] Preferably, in step 5, the solution formula of the class discrimination saliency map is as follows:

[0028]

[0029] Among them, L c is the class discrimination saliency map of class c strata, represents the weight corresponding to the k-th channel feature map, where The solution formula is as follows:

[0030]

[0031] in, is the global confidence score.

[0032] A stratum partition model visualization system using confidence scores, used to implement the above-mentioned stratum partition model visualization method using confidence scores, comprising:

[0033] The drawing module is used to draw the spatial position coordinates of all oil wells in the block and select the oil wells with similar spatial position coordinates as a data set, which includes several logging curve data;

[0034] The first data processing module is used to remove abnormal values ​​of the logging curve data in the data set, standardize the logging curve data of each oil well according to the curve, select a sliding window to obtain a two-dimensional representation segment of the two-dimensional logging data, convert all the logging curve data in the form of discrete data in each well into two-dimensional segments of the size of the sliding window, obtain the formation label corresponding to each segment, vectorize all geological layer labels to obtain a data set, and divide the obtained data set into a training set and a test set according to the ratio;

[0035] A model construction module, which is used to construct a convolutional neural network model, input the data in the training set into the convolutional neural network to obtain the output result of the network, compare the output result of the network with the formation label, calculate the loss function, and then calculate the gradient of the network parameters with respect to the loss function through the backpropagation algorithm. Repeat the training process until the preset maximum number of training epochs is reached to complete the training;

[0036] A second data processing module, which is used to input the two-dimensional characterization segment of the well logging data to be visualized into the trained convolutional neural network, select the convolutional layer that needs to be visualized, and output the feature map of this layer; after upsampling the output feature map to the size of the input segment and multiplying it point by point with the two-dimensional characterization segment of the input well logging data, re-input it into the convolutional neural network with the same structure to obtain the confidence score of each feature map;

[0037] A fourth data processing module, which is used to linearly weight the confidence scores of the feature maps of each channel with the feature maps of the corresponding channels to obtain a class discrimination saliency map;

[0038] An execution module, which is used to repeat the execution to visualize multiple two-dimensional characterization segments of well logging data of the same formation and obtain the corresponding class discrimination saliency map.

[0039] A mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for visualizing a formation division model using confidence scores are implemented.

[0040] Compared with the prior art, the present invention has the following beneficial technical effects:

[0041] The present invention provides a visualization method for a formation division model using confidence scores. First, by plotting the spatial position coordinates of oil wells and screening for similar oil wells, a dataset is formed. Then, outlier removal and normalization processing are performed on the logging curves in the dataset, and two-dimensional characterization segments are generated using a sliding window. At the same time, the formation labels are vectorized. Next, a convolutional neural network model is constructed and the model is trained using the training set data. Through multiple iterations, the network parameters are optimized so that the network can accurately predict the formation labels. After training is completed, the logging data segments to be visualized are input into the trained network, and feature maps are output at the specified convolutional layer. Then, the feature maps are upsampled and multiplied point by point with the input segments, and then input into a convolutional neural network with the same structure again to obtain the confidence scores of the feature maps. The confidence scores of each channel feature map are used as weights to linearly weight the corresponding channel feature maps to obtain a class discrimination saliency map. Finally, the above technology is used to visualize multiple logging data segments of the same formation to obtain a class discrimination saliency map, thereby revealing the electrical properties of the formation and its mapping results in the feature space. The visualization method for the formation division model using confidence scores gets rid of the dependence on gradients during backpropagation of the convolutional neural network, and can visualize any convolutional layer in the network, which helps to understand the electrical properties of the formation and provides important references in geological research and oil exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a schematic flow chart of the visualization method for the formation division model using confidence scores in the present invention;

[0043] Figure 2 is a geographical location coordinate map of all wells in a certain block of a certain oilfield in an embodiment of the present invention;

[0044] Figure 3 is a geographical location coordinate map of the dataset in the present invention;

[0045] Figure 4 is a schematic process diagram of converting logging curve data into two-dimensional characterization segments in the present invention;

[0046] Figure 5 is the convolutional neural network structure diagram constructed in the present invention;

[0047] Figure 6 is a schematic diagram of the visualization mechanism of the formation division model using confidence scores in the present invention;

[0048] Figure 7 is the visualization result map of the Chang6 geological formation in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme 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 only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0050] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0051] The purpose of the present invention is to provide a method, system and terminal for visualizing a stratigraphic division model using confidence scores, so as to solve the technical problems that the main manual interpretation method for stratigraphic division in the prior art has too many subjective factors and a large number of errors caused by human factors.

[0052] See also Figure 1 The present invention provides a method for visualizing a stratigraphic partitioning model using confidence scores, comprising the following steps:

[0053] Step 1, draw the spatial position coordinates of all oil wells in the block, and select the oil wells with similar spatial position coordinates as a data set, the data set includes several well logging curve data;

[0054] Among them, the hierarchical clustering method is used to preferentially screen oil wells with similar spatial coordinates as the data set. Several logging curve data in the data set correspond to the data sources of several features. The logging curve data are presented in the form of discrete data of a single depth point, and each depth point corresponds to a formation attribute.

[0055] Among them, several logging curve data include depth data, natural gamma ray GR data, natural potential SP data, acoustic wave time difference AC data, array induction resistivity AT90 data and array induction resistivity AT20 data.

[0056] Step 2: remove outliers in the well logging curve data in the data set, standardize the well logging curve data of each oil well according to the curve, select a sliding window to obtain a two-dimensional representation fragment of the two-dimensional well logging data, convert all the well logging curve data in the form of discrete data in each well into two-dimensional fragments of the sliding window size, obtain the formation label corresponding to each fragment, vectorize all geological layer labels to obtain a data set, and divide the obtained data set into a training set and a test set according to the ratio;

[0057] Among them, the logging curve data of each oil well is standardized according to the curve, and the standardization is Z-Score standardization, and the formula is as follows:

[0058]

[0059] Among them, μ is the data mean value, and σ is the data standard deviation.

[0060] Among them, the size of the sliding window is selected as 96×6 pixels, and it slides along the direction of the well logging curve data at an interval of 1 depth point. Each time, a two-dimensional characterization segment of well logging data with a size of 96×6 can be obtained within the window coverage; all the well logging curve data in the form of discrete data in each well are converted into two-dimensional segments with a size of 96×6. The formation label corresponding to each segment is the formation label that appears the most times among the 96 depth points selected by the window. Finally, all the geological layer labels are vectorized. The specific process of vectorizing all the geological layer labels is as follows:

[0061] Count the types n of formation labels in the training set, number the formation labels from 0, and then vectorize the numbers of each label into n-dimensional vectors, where one index in the vector is 1 and the rest of the indices are 0.

[0062] Step 3, construct a convolutional neural network model. Input the data in the training set into the convolutional neural network to obtain the output result of the network. Compare the output result of the network with the formation label, calculate the loss function, and then calculate the gradient of the network parameters with respect to the loss function through the backpropagation algorithm. Repeat the training process until the preset maximum number of training rounds is reached to complete the training;

[0063] Specifically, first design the architecture of the convolutional neural network in constructing the convolutional neural network model. The architecture of the convolutional neural network includes 3 convolutional layers, 2 fully connected layers, and a softmax classifier. The number of convolutional kernels in the 3 convolutional layers is 64, 128, and 256 respectively. The size of the convolutional kernel is 3×3, the stride is 1, and the padding method is same; the same padding method means padding a circle of 0s around the array to increase the size of the array; the number of nodes in both fully connected layers is 512; the number of nodes in the softmax classifier is determined by the counted number of formation label types.

[0064] Step 4, input the two-dimensional characterization segment of the well logging data to be visualized into the trained convolutional neural network, select the convolutional layer that needs to be visualized, and output the feature map of this layer; upsample the output feature map to the size of the input segment and multiply it point by point with the two-dimensional characterization segment of the input well logging data, and then re-enter it into the convolutional neural network with the same structure to obtain the confidence score of each feature map;

[0065] Specifically, the formula for solving the confidence score of the two-dimensional characterization segment to be visualized is as follows:

[0066]

[0067] Among them, is the output feature map of the k-th channel of the l-th convolutional layer, is the global confidence score for the two-dimensional representation segment scores;

[0068] The calculation formula of is:

[0069] where the Up function means upsampling the feature map to the dimension of the input vector, and the Norm function means normalizing the upsampled feature map.

[0070] Step 5, use the confidence score of each channel feature map as the weight to linearly weight with the corresponding channel feature map to obtain the class discriminant saliency map;

[0071] where the solution formula of the class discriminant saliency map is as follows:

[0072]

[0073] where L c is the class discriminant saliency map of the c-th class formation, represents the weight corresponding to the k-th channel feature map, where The solution formula of is as follows:

[0074]

[0075] where, is the global confidence score.

[0076] Step 6, repeat Steps 1 to 5, visualize the two-dimensional representation segments of multiple logging data of the same formation, obtain the corresponding class discriminant saliency maps, analyze the common patterns presented by them, and summarize the mapping of the electrical characteristics of the formation in the feature space. The results show that the deep features of the logging curves are mainly the curve features in the middle area, and the interaction of the curves shows an obvious periodic vertical arrangement. It can be seen from the visualization result map that there is a phenomenon of uneven color in the areas of interest of the network, indicating that different logging curves have different contributions to the network decision-making.

[0077] Embodiment

[0078] This embodiment provides a visualization method for a formation division model using confidence scores, and the specific process is as follows:

[0079] Step 1, block data screening.

[0080] In this embodiment, first, a geographical location coordinate map of all wells in a certain block of a certain oilfield is drawn, such as Figure 2As shown, 228 wells with similar geographical locations were selected as the dataset through hierarchical clustering. 28 wells were randomly selected as test wells, and the remaining 200 wells were training wells. The well location map of the dataset is as shown in Figure 3 As shown. Each well in the dataset contains 6 logging curve data, namely Depth, Gamma Ray (GR), Spontaneous Potential (SP), Acoustic Travel Time (AC), Array Induction Resistivity AT90, and Array Induction Resistivity AT20. The logging curve data of each well is presented in the form of discrete data at a single depth point. Each depth point corresponds to a geological horizon, and there are 10 types of geological horizons, namely 'K1z2+1', 'J2z', 'J1y', 'J2a', 'J1f', 'Chang1', 'Chang2', 'Chang3', 'Chang4+5', 'Chang6'. The form of the dataset is shown in Table 1.

[0081] Table 1 Example of the dataset form

[0082] Well Name Depth(m) GR(gAPI) SP(mv) AC(us / m) AT90(Ω·m) AT20(Ω·m) Label W100 648.000 50.83 40.87 263.05 14.93 15.40 J1a W100 648.125 53.03 41.73 269.79 13.44 14.54 J1a … … … … … … … … W100 1270.000 97.21 66.91 243.65 18.82 20.33 J1f W100 1270.125 96.60 66.73 244.11 15.71 17.84 Chang1 … … … … … … … … W100 1807.625 0.00 78.83 239.94 21.72 8.12 Chang6 W100 1807.750 0.00 78.70 237.83 21.47 8.08 Chang6

[0083] Step 2: Two-dimensional characterization of logging data.

[0084] First, outliers such as 0, -1, and 9999 generated due to logging sensor jitter and other reasons in the logging curve data of each well in the dataset were removed to prevent outliers from affecting the network training effect. Then, the mean and standard deviation of each logging curve data of each well were calculated. According to the standardization formula:

[0085] The logging curve data of each well was standardized by curve to avoid excessive data deviation. Then, a sliding window of size 96×6 was selected and slid along the direction of the logging curve data at an interval of 1 depth point. Each time, a two-dimensional logging data two-dimensional characterization segment of size 96×6 could be obtained within the window coverage. The process of converting the logging curve data into the form of two-dimensional characterization segments is as shown in Figure 4 As shown. Finally, all 6 logging curve data in the form of discrete data in each well were converted into two-dimensional characterization segments of logging data of size 96×6. The formation label corresponding to each segment was the formation label that appeared most frequently among the 96 depth points selected by the window. Finally, all geological formation labels were vectorized. That is, the types of geological formation labels in the training set were counted as 10 types, numbered from 0 for the geological formation labels, and then the numbers of each label were vectorized into 10-dimensional vectors, where one index in the vector was 1 and the rest were 0.

[0086] Step 3: Build and train a convolutional neural network model. First, a convolutional neural network containing 3 convolutional layers, 2 fully connected layers, and a softmax classifier was built for formation classification. The structure of the convolutional neural network is as shown in Figure 5As shown in the figure, the number of convolution kernels in the three convolution layers of the convolutional neural network is 64, 128, and 256 respectively. The size of the convolution kernel is 3×3, the stride is 1, and the padding method is same padding, that is, a circle of 0s is filled around the array to increase the size of the array. During convolution, the convolution kernel can exceed the boundary of the original array for convolution, preventing the loss of edge information during the convolution process and ensuring that the size of the result after convolution is the same as that before convolution. A batch normalization layer is added between each convolution layer to prevent overfitting during network training. The number of nodes in both fully connected layers is 512. Since there are 10 types of geological layers in total, the number of nodes in the softmax classifier is set to 10. The activation function of the convolution layer and the fully connected layer uses ReLU, and the network training optimization algorithm uses the Adam algorithm. Input the two-dimensional logging data characterization segment after preprocessing in step two into the convolutional neural network, update the network parameters according to the formation labels corresponding to the two-dimensional segments, and finally train to obtain the network model and save it;

[0087] Step four, solve the confidence score of the two-dimensional characterization segment to be visualized. Input the two-dimensional characterization segment of the logging data to be visualized into the trained convolutional neural network, select the convolution layer that needs to be visualized, and output the feature map of this layer. After upsampling the output feature map to the size of the input segment and multiplying it point by point with the two-dimensional characterization segment of the input logging data, input it back into the convolutional neural network with the same structure to obtain the confidence score of each feature map. The formula is:

[0088]

[0089] Among them, is the output feature map of the kth channel of the lth convolution layer, is the global confidence score for the score of the two-dimensional characterization segment, The calculation formula of

[0090]

[0091] Among them, the Up function represents upsampling the feature map to the dimension of the input vector, and the Norm function represents normalizing the upsampled feature map.

[0092] Step five, visualize the two-dimensional characterization segment. Regard the confidence score of each channel feature map obtained in step four as the weight and linearly weight it with the corresponding channel feature map to obtain the class discriminant saliency map, realizing the visualization of the two-dimensional characterization segment. Its visualization mechanism is as Figure 6 shown. The formula for solving the class discriminant saliency map of the corresponding segment is:

[0093]

[0094] Among them, L cIt is a significant map for class discrimination of formation C. It represents the weight corresponding to the k-th channel feature map, and its solution formula is:

[0095]

[0096] Where is the global confidence score obtained in Step 4.

[0097] Step 6: Analyze and summarize the mapping law of formation electrical characteristics in the feature space. Visualize the two-dimensional characterization segments of multiple logging data of the same formation to obtain the corresponding significant map for class discrimination, analyze the common laws presented, and summarize the mapping of the electrical characteristics of this formation in the feature space. Considering that the Chang 6 layer in the formation has the most oil and gas content, the present invention mainly visualizes the two-dimensional characterization segments of logging data of the Chang 6 geological layer. The visualization results are as Figure 7 shown. Figure 7 lists the visualization results of three wells, namely W615, W189, and W634, in the test set. For each well, two two-dimensional characterization segments belonging to the Chang 6 layer are randomly selected for visualization. Considering the professionalism of the logging field, in this paper, when presenting the visualization results, they are fed back to the logging curve level for comprehensive analysis. On the leftmost side of each figure is the curve shape of five logging curves, namely GR, SP, AC, AT90, and AT20, after standardization. The ordinate is the depth of the sampling points. To achieve the visibility of the curves, a 36-meter logging curve segment is drawn for each figure, and the ordinate depth interval is 12 meters. Each interval corresponds to a two-dimensional characterization segment of logging data. On the right is the visualization result of the Chang 6 layer data. The non-blue areas in the figure are the deep features of the Chang 6 layer extracted by the network. The redder the color, the greater the positive impact of this area on the network decision. The rightmost blue ellipse in each figure is a local enlarged view of the deep features. It can be seen from the figure that the deep features of the logging curves of the Chang 6 layer are mainly the curve features in the middle area, and the curves interact to show an obvious periodic vertical arrangement. From the visualization result figure, it can be seen that there is a phenomenon of uneven color in the areas of interest to the network, indicating that different logging curves have different contributions to the network decision.

[0098] In this embodiment, the visualization method of the formation division model using the confidence score is used to visualize the Chang 6 geological layer with the most oil and gas content, and the mapping law of its electrical characteristics in the feature space can be clearly found, which shows the effectiveness of the present invention and the feasibility of this method.

[0099] The present invention also provides a stratum division model visualization system using confidence scores, which is used to implement the above-mentioned stratum division model visualization method using confidence scores, including a drawing module, a first data processing module, a model building module, a second data processing module, a fourth data processing module and an execution module;

[0100] The drawing module is used to draw the spatial position coordinates of all oil wells in the block and select the oil wells with similar spatial position coordinates as a data set, which includes several logging curve data;

[0101] The first data processing module is used to remove abnormal values ​​of the logging curve data in the data set, standardize the logging curve data of each oil well according to the curve, select a sliding window to obtain a two-dimensional representation segment of the two-dimensional logging data, convert all the logging curve data in the form of discrete data in each well into two-dimensional segments of the size of the sliding window, obtain the formation label corresponding to each segment, vectorize all geological layer labels to obtain a data set, and divide the obtained data set into a training set and a test set according to the ratio;

[0102] The model building module is used to build a convolutional neural network model, input the data in the training set into the convolutional neural network to obtain the network output, compare the network output with the formation label, and calculate the loss function. Then, the gradient of the network parameters to the loss function is calculated through the back propagation algorithm, and the training process is repeated until the preset maximum number of training rounds is reached to complete the training;

[0103] The second data processing module is used to input the two-dimensional representation fragment of the well logging data to be visualized into the trained convolutional neural network, select the convolution layer to be visualized, and output the feature map of the layer; upsample the output feature map to the input fragment size and multiply it point by point with the input two-dimensional representation fragment of the well logging data, and then re-input it into the convolutional neural network with the same structure to obtain the confidence score of each feature map;

[0104] A fourth data processing module is used to use the confidence score of each channel feature map as a weight and linearly weight the feature map of the corresponding channel to obtain a class discrimination saliency map;

[0105] The execution module is used for repeated execution to visualize the two-dimensional representation fragments of multiple well logging data of the same formation and obtain the corresponding class discrimination saliency map.

[0106] The present invention also provides a mobile terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, such as a visualization program for a stratigraphic division model using confidence scores.

[0107] When the processor executes the computer program, it implements the steps of the above-mentioned formation division model visualization method using confidence scores, including the following steps:

[0108] Step 1, plot the spatial position coordinates of all oil wells in the block, and screen out the oil wells with similar spatial position coordinates as the data set, where the data set includes several logging curve data;

[0109] Step 2, remove the outliers in the logging curve data of the data set, standardize the logging curve data of each oil well according to the curve, select a sliding window to obtain the two-dimensional representation segments of the two-dimensional logging data, convert all the logging curve data in the form of discrete data in each well into two-dimensional segments of the sliding window size, obtain the formation labels corresponding to each segment, vectorize all the geological layer labels to obtain the data set, and divide the obtained data set into a training set and a test set according to a ratio;

[0110] Step 3, construct a convolutional neural network model, input the data in the training set into the convolutional neural network to obtain the output result of the network, compare the output result of the network with the formation labels, calculate the loss function, and then calculate the gradient of the network parameters with respect to the loss function through the backpropagation algorithm. Repeat the training process until the preset maximum number of training epochs is reached to complete the training;

[0111] Step 4, input the two-dimensional representation segments of the logging data to be visualized into the trained convolutional neural network, select the convolutional layer that needs to be visualized, and output the feature map of this layer; after upsampling the output feature map to the size of the input segment and multiplying it point by point with the two-dimensional representation segment of the input logging data, re-enter it into the convolutional neural network with the same structure to obtain the confidence score of each feature map;

[0112] Step 5, use the confidence scores of each channel feature map as weights to linearly weight the feature map of the corresponding channel to obtain the class discriminant saliency map;

[0113] Step 6, repeat Steps 1 to 5, visualize the two-dimensional representation segments of multiple logging data of the same formation, obtain the corresponding class discriminant saliency map and obtain the mapping result of the electrical characteristics of this formation in the feature space.

[0114] Alternatively, when the processor executes the computer program, it implements the functions of each module in the above system. For example, a drawing module is used to plot the spatial position coordinates of all oil wells in the block, and screen out the oil wells with similar spatial position coordinates as the data set, where the data set includes several logging curve data;

[0115] The first data processing module is used to remove abnormal values ​​of the logging curve data in the data set, standardize the logging curve data of each oil well according to the curve, select a sliding window to obtain a two-dimensional representation segment of the two-dimensional logging data, convert all the logging curve data in the form of discrete data in each well into two-dimensional segments of the size of the sliding window, obtain the formation label corresponding to each segment, vectorize all geological layer labels to obtain a data set, and divide the obtained data set into a training set and a test set according to the ratio;

[0116] The model building module is used to build a convolutional neural network model, input the data in the training set into the convolutional neural network to obtain the network output, compare the network output with the formation label, and calculate the loss function. Then, the gradient of the network parameters to the loss function is calculated through the back propagation algorithm, and the training process is repeated until the preset maximum number of training rounds is reached to complete the training;

[0117] The second data processing module is used to input the two-dimensional representation fragment of the well logging data to be visualized into the trained convolutional neural network, select the convolution layer to be visualized, and output the feature map of the layer; upsample the output feature map to the input fragment size and multiply it point by point with the input two-dimensional representation fragment of the well logging data, and then re-input it into the convolutional neural network with the same structure to obtain the confidence score of each feature map;

[0118] A fourth data processing module is used to use the confidence score of each channel feature map as a weight and linearly weight the feature map of the corresponding channel to obtain a class discrimination saliency map;

[0119] The execution module is used for repeated execution to visualize the two-dimensional representation fragments of multiple well logging data of the same formation and obtain the corresponding class discrimination saliency map.

[0120] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program in the mobile terminal. For example, the computer program may be divided into a drawing module, a first data processing module, a model building module, a second data processing module, a fourth data processing module, and an execution module;

[0121] The specific functions of each module are as follows:

[0122] The drawing module is used to draw the spatial position coordinates of all oil wells in the block and select the oil wells with similar spatial position coordinates as a data set, which includes several logging curve data;

[0123] The first data processing module is used to remove abnormal values ​​of the logging curve data in the data set, standardize the logging curve data of each oil well according to the curve, select a sliding window to obtain a two-dimensional representation segment of the two-dimensional logging data, convert all the logging curve data in the form of discrete data in each well into two-dimensional segments of the size of the sliding window, obtain the formation label corresponding to each segment, vectorize all geological layer labels to obtain a data set, and divide the obtained data set into a training set and a test set according to the ratio;

[0124] The model building module is used to build a convolutional neural network model, input the data in the training set into the convolutional neural network to obtain the network output, compare the network output with the formation label, and calculate the loss function. Then, the gradient of the network parameters to the loss function is calculated through the back propagation algorithm, and the training process is repeated until the preset maximum number of training rounds is reached to complete the training;

[0125] The second data processing module is used to input the two-dimensional representation fragment of the well logging data to be visualized into the trained convolutional neural network, select the convolution layer to be visualized, and output the feature map of the layer; upsample the output feature map to the input fragment size and multiply it point by point with the input two-dimensional representation fragment of the well logging data, and then re-input it into the convolutional neural network with the same structure to obtain the confidence score of each feature map;

[0126] A fourth data processing module is used to use the confidence score of each channel feature map as a weight and linearly weight the feature map of the corresponding channel to obtain a class discrimination saliency map;

[0127] The execution module is used for repeated execution to visualize the two-dimensional representation fragments of multiple well logging data of the same formation and obtain the corresponding class discrimination saliency map.

[0128] The mobile terminal may be a computing device such as a desktop computer, a notebook, a palm computer, a cloud server, etc. The mobile terminal may include, but is not limited to, a processor and a memory.

[0129] The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the mobile terminal, and connects various parts of the entire mobile terminal using various interfaces and lines.

[0130] The memory can be used to store the computer program and / or module. By running or executing the computer program and / or module stored in the memory, and by calling the data stored in the memory, the processor realizes various functions of the mobile terminal.

[0131] The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage devices.

[0132] Finally, it should be noted that: The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: It is still possible to modify the specific implementation manners of the present invention or make equivalent replacements. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A visualization method for stratigraphic partitioning model using confidence scores, It is characterized in that The steps include: Step 1, draw the spatial position coordinates of all oil wells in the block, and select the oil wells with similar spatial position coordinates as a data set, the data set includes several logging curve data; Step 2: remove outliers in the well logging curve data in the data set, standardize the well logging curve data of each oil well according to the curve, select a sliding window to obtain a two-dimensional representation fragment of the two-dimensional well logging data, convert all the well logging curve data in the form of discrete data in each well into two-dimensional fragments of the sliding window size, obtain the formation label corresponding to each fragment, vectorize all geological layer labels to obtain a data set, and divide the obtained data set into a training set and a test set according to the ratio; Step 3: Construct a convolutional neural network model, input the data in the training set into the convolutional neural network to obtain the network output, compare the network output with the formation label, and calculate the loss function. Then, calculate the gradient of the network parameters to the loss function through the back propagation algorithm, and repeat the training process until the preset maximum number of training rounds is reached to complete the training. Step 4: Input the two-dimensional representation fragment of the well logging data to be visualized into the trained convolutional neural network, select the convolutional layer to be visualized, and output the feature map of the layer; upsample the output feature map to the input fragment size and multiply it point by point with the input two-dimensional representation fragment of the well logging data, and then re-input it into the convolutional neural network with the same structure to obtain the confidence score of each feature map; Step 5: Use the confidence score of each channel feature map as a weight and linearly weight it with the feature map of the corresponding channel to obtain a class discrimination saliency map; Step 6, repeating steps 1 to 5, visualizing multiple two-dimensional characterization segments of logging data of the same formation, obtaining corresponding class discrimination saliency maps and obtaining mapping results of the electrical characteristics of the formation in the feature space.

2. A method for visualizing a stratigraphic partitioning model using confidence scores according to claim 1, It is characterized in that In step 1, the hierarchical clustering method is used to preferentially select oil wells with similar spatial coordinates as the data set. Several logging curve data in the data set correspond to the data sources of several features, where the logging curve data are presented in the form of discrete data of a single depth point, and each depth point corresponds to a formation attribute.

3. The method for visualizing a stratigraphic division model using confidence scores according to claim 1, It is characterized in that Several logging curve data include depth data, natural gamma ray GR data, natural potential SP data, acoustic wave time difference AC data, array induction resistivity AT90 data and array induction resistivity AT20 data.

4. The method for visualizing a stratigraphic partitioning model using confidence scores according to claim 1, It is characterized in that In step 2, the logging curve data of each oil well is standardized according to the curve, and the standardization is Z-Score standardization, and the formula is as follows: Among them, μ is the data mean and σ is the data standard deviation.

5. The method for visualizing a stratigraphic partitioning model using confidence scores according to claim 1, It is characterized in that In step 2, the size of the sliding window is selected as 96×6 pixels, and it slides along the direction of the well logging curve data at intervals of 1 depth point. Each time within the window coverage, a 96×6 2D well logging data 2D representation fragment can be obtained; several well logging curve data in the form of discrete data in each well are all converted into 96×6 2D fragments, where the formation label corresponding to each fragment is the formation label that appears most frequently in the 96 depth points selected by the window. Finally, all geological layer labels are vectorized, and the specific process of vectorizing all geological layer labels is as follows: Count the types of stratigraphic labels in the training set, number the stratigraphic labels starting from 0, and then vectorize the number of each label into an n-dimensional vector, where one index in the vector is 1 and the rest are 0.

6. The method for visualizing a stratigraphic partitioning model using confidence scores according to claim 1, It is characterized in that In step 3, the architecture of the convolutional neural network is first designed in constructing the convolutional neural network model. The architecture of the convolutional neural network includes 3 convolutional layers, 2 fully connected layers and a softmax classifier, wherein the number of convolution kernels of the 3 convolutional layers are 64, 128 and 256 respectively, the convolution kernel size is 3×3, the step size is 1, and the filling method is same; the same filling method is to fill a circle of 0 around the array to increase the array size; the number of nodes in the two fully connected layers is 512; the number of nodes of the softmax classifier is determined by the number of statistical stratum label types.

7. The method for visualizing a stratigraphic partitioning model using confidence scores according to claim 1, It is characterized in that In step 4, the confidence score of the two-dimensional representation fragment to be visualized is calculated as follows: wherein, is the output feature map of the k-th channel of the l-th convolutional layer, is the global confidence score for the two-dimensional characterization segment score; The calculation formula is as follows: Among them, the Up function means upsampling the feature map to the dimension of the input vector, and the Norm function means normalizing the upsampled feature map.

8. The method for visualizing a stratigraphic partitioning model using confidence scores according to claim 1, It is characterized in that In step 5, the solution formula for the class discrimination saliency map is as follows: Among them, L c is the class discrimination significant map of the C-type formation, represents the weight corresponding to the k-th channel feature map, where The solution formula of is as follows: Among them, is the global confidence score.

9. A system for visualizing a stratigraphic division model using confidence scores, used to implement a method for visualizing a stratigraphic division model using confidence scores as claimed in any one of claims 1 to 8, It is characterized in that include: The drawing module is used to draw the spatial position coordinates of all oil wells in the block and select the oil wells with similar spatial position coordinates as a data set, which includes several logging curve data; The first data processing module is used to remove abnormal values ​​of the logging curve data in the data set, standardize the logging curve data of each oil well according to the curve, select a sliding window to obtain a two-dimensional representation segment of the two-dimensional logging data, convert all the logging curve data in the form of discrete data in each well into two-dimensional segments of the size of the sliding window, obtain the formation label corresponding to each segment, vectorize all geological layer labels to obtain a data set, and divide the obtained data set into a training set and a test set according to the ratio; A model construction module, configured to construct a convolutional neural network model, input data in a training set into the convolutional neural network to obtain an output result of the network, compare the output result of the network with a formation label, calculate a loss function, and then calculate the gradient of the network parameters with respect to the loss function through a backpropagation algorithm, and repeat the training process until a preset maximum number of training rounds is reached to complete the training; A second data processing module, configured to input a two-dimensional characterization segment of well logging data to be visualized into the trained convolutional neural network, select a convolutional layer that needs to be visualized, and output a feature map of this layer; after upsampling the output feature map to the size of the input segment and multiplying it point by point with the two-dimensional characterization segment of the input well logging data, re-input it into the convolutional neural network with the same structure to obtain a confidence score for each feature map; A fourth data processing module, configured to use the confidence score of each channel feature map as a weight to linearly weight the feature map of the corresponding channel to obtain a class discrimination saliency map; An execution module, configured to repeat the execution to visualize multiple two-dimensional characterization segments of well logging data of the same formation, and obtain corresponding class discrimination saliency maps.

10. A mobile terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, the steps of a method for visualizing a formation division model using confidence scores according to any one of claims 1 to 8 are implemented.

Images