A phase determination method based on sandstone section and formation section logging curves
By constructing a multi-input deep learning model and combining well logging curve data from sandstone and stratigraphic sections, the problem of complex and insufficient accuracy in sedimentary facies division in existing technologies has been solved, achieving high-precision sedimentary facies identification.
Patent Information
- Application Number
- CN202510596414.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-09
- Publication Date
- 2026-11-10
AI Technical Summary
Existing methods for classifying sedimentary facies are complex and lack precision, while deep learning models have low accuracy in facies identification.
A multi-input deep learning model based on logging curves of sandstone and formation sections is adopted. By constructing multiple input branches, phase determination is performed by combining sandstone parameters and various logging curve data, including data cleaning, normalization processing and multi-input model training.
It improves the accuracy and efficiency of sedimentary facies identification, with model training accuracy reaching over 90%.
Smart Images

Figure CN123069008A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fine reservoir characterization technology, and in particular to a phase determination method based on logging curves of sandstone and formation sections. Background Technology
[0002] Sedimentary facies, as a crucial research area in fine reservoir characterization of oilfields, serves as a vital geological basis for calculating geological reserves and predicting production in developed blocks. Traditional methods for sedimentary facies classification generally rely on the regional sedimentary background, combined with various well logging curve morphological characteristics, and comprehensive analysis of rock thickness, grain size, and other features to delineate sedimentary microfacies in the study area. However, this process requires the comprehensive analysis of multiple datasets, making it complex, tedious, and challenging. With the development of machine learning technology, deep learning-based facies identification methods based on well logging data have emerged. However, these methods can only perform deep learning on a single well logging curve, and the highest accuracy achieved after training is only around 85%, which is insufficient. To address these shortcomings, a facies identification method based on well logging curves from sandstone and stratigraphic sections is proposed. Summary of the Invention
[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a phase determination method based on well logging curves of sandstone and formation sections, which solves the problems of complex and difficult manual phase determination processes and low accuracy of phase determination using deep learning models in existing technologies.
[0004] (II) Technical Solution To address the above problems, this invention provides a phase determination method based on well logging curves of sandstone and formation sections, comprising: Step S1: Collect formation and production well data within the target area; Step S2: Analyze and clean the data collected in step S1 to eliminate outliers; Step S3: Create multiple data samples X1, X2...Xn based on the data processed in step S2, and construct a dataset {X1i, X2i, ..., Xni} based on the established samples. At least one sample in the data samples is a sandstone parameter sample. Step S4: Construct a multi-input model. The multi-input model has multiple input branches at the input end. The number of input branches corresponds to the number of samples. Within each input branch, select the sub-learning model corresponding to the input data, and merge the output results of each input end to output the classification result. Step S5: Input the dataset obtained in step S3 into the multi-input model constructed in step S4 for training; Step S6: Evaluate the training results of the multi-input model; Step S7: Input the data that needs to be phased into the trained multi-input model for predictive analysis.
[0005] Furthermore, in step S1, the sandstone data table in the target area is expanded based on the collected data, adding fields for bottom depth and thickness of the formation.
[0006] Furthermore, in step S2, data with invalid and missing values are cleaned up, and then the remaining data is normalized.
[0007] Furthermore, during the normalization process, the logging curve is scaled to a range of {0-255} pixels.
[0008] Furthermore, the logging curves include micro-gradient curves, micro-potential curves, natural gamma curves, and natural potential curves.
[0009] Furthermore, the data samples in step S3 include the X1 sandstone section logging curve, the X2 sandstone parameters, and the X3 formation section logging curve.
[0010] Furthermore, after the samples are constructed in step S3, the obtained samples are classified according to sedimentary facies and the number of samples of each type of stratum is counted.
[0011] Furthermore, in step S4, the sub-learning models include CNN models and RNN models. After the input of the multi-input model, the output results of each sub-learning model at the input end are merged by concatenation technology to output the calculation result.
[0012] Furthermore, in step S5, the sample dataset obtained in step S3 is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. During the training process, the content of the training set is substituted into the multi-input model for training and learning, and the learning results of the multi-input model are verified and corrected through the validation set.
[0013] Furthermore, in step S6, the test set is input into the trained multi-input model, and the output of the multi-input model is compared with the actual result. The training effect of the multi-input model is judged based on the accuracy between the two.
[0014] (III) Beneficial Effects The facies determination method based on well logging curves of sandstone and stratigraphic sections provided by this invention establishes a multi-input depth model sample. Multiple input branches are set up at the input end of the multi-input model, and each branch learns from a data sample. The output results of each input branch are combined to determine the facies. At the same time, the influence of multiple samples on the facies determination results is learned, and multiple data such as sandstone parameters are taken into account when determining sedimentary facies. It does not only learn from a single data, but also incorporates more data information in the facies determination process, thereby improving the accuracy of the learning model in determining sedimentary facies. Attached Figure Description
[0015] Figure 1 This is a flowchart of the phase determination method based on well logging curves of sandstone and formation sections according to the present invention; Figure 2 This is an example diagram of sandstone data in the facies determination method based on well logging curves of sandstone and formation sections according to the present invention; Figure 3 This is an example diagram of the data distribution after processing by the phase determination method based on well logging curves of sandstone and formation sections according to the present invention; Figure 4 This is a normalized example of the logging curves for the phase determination method based on logging curves of sandstone and formation sections according to the present invention; Figure 5 This is a diagram illustrating the RNN model architecture of the phase determination method based on well logging curves of sandstone and formation sections in this invention. Figure 6 This is a diagram illustrating the CNN model architecture of the phase determination method based on well logging curves of sandstone and formation sections in this invention. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] In the description of this invention, it is necessary to understand that the orientation or positional relationship indicated by terms such as "upper", "lower", "inner", "outer", "top", and "bottom" are based on the orientation or positional relationship shown in the accompanying drawings. The purpose is only to facilitate the description of this invention and to simplify the description. It is not intended to indicate or imply that the component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention. Example
[0018] like Figure 1 As shown, this invention provides a phase determination method based on well logging curves of sandstone and formation sections, specifically including: Step S1: Collect formation and production well data within the target area. Collect logging curves, formation data tables, and sandstone data tables from each well within the target area as data sources for subsequent experiments. These logging curves, formation data tables, and sandstone data tables are obtained during the production and logging processes of each production well. During data acquisition, the tables and charts can be directly retrieved from the historical construction records of each well. For some production wells with longer development histories or earlier development periods, data may be missing. In such cases, logging operations can be performed separately on the production wells with missing data to obtain the necessary logging data.
[0019] After collecting the required formation and production well data, the sandstone data table for the target area is expanded by adding fields for formation depth and formation thickness. After adding the necessary fields to the sandstone data table, each data point in the sandstone data table is correlated with its formation range, facilitating subsequent data classification based on formation.
[0020] Step S2: Analyze and clean the collected data to eliminate outliers. After data collection, during the data acquisition, storage, and transmission process, data distortion and loss are inevitable due to signal loss or other reasons. In order to ensure the accuracy of subsequent work, it is necessary to process the outliers caused by distortion and loss in the collected data.
[0021] This process involves cleaning up data containing invalid and missing values, and then normalizing the remaining data. Typically, the collected downhole data should follow a normal distribution. According to the Iradha principle, a data point value is considered an outlier if its deviation from the mean exceeds three times the standard deviation. The calculation formula is as follows: |x−μ|>3σ, where x is the data value, μ is the mean of the data, and σ is the standard deviation of the data. Typically, the probability of outliers in a dataset should be below 0.3%, i.e. P(|x−μ|>3σ)≤0.003.
[0022] In addition, for some well logging curves in the collected data that have missing or incomplete data, in order to ensure the accuracy of subsequent multi-input model training, well logging curves with outliers and missing values and their corresponding well logging data are usually deleted directly.
[0023] It is important to note that the normalization process scales the logging curves to a range of {0-255} pixels. First, the collected data is uniformly standardized using Z-score normalization to convert the logging curves into a data distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the influence of dimensions in the data and improving the accuracy of subsequent model training. The calculation formula is as follows: Z = (X-μ) / σ, where Z is the converted value, X is the logging value before conversion, μ is the mean of the dataset, and σ is the standard deviation of the dataset.
[0024] After the first standardization process, for the well logging curves in the collected data, in order to directly input the well logging curves as input data to train the multi-input model, it is necessary to convert the well logging curves into a pixel value range between 0 and 255. However, directly converting the well logging curve data that has only undergone Z-score standardization into an image with a pixel value range of 0-255 may cause some data to fail to meet the criteria for judging outliers. Therefore, a second normalization transformation is required to scale the well logging curve data to a fixed range of 0-1 to ensure the legality and accuracy of image pixels in subsequent transformation processes.
[0025] Typically, well logging curves include microgradient curves, micropotential curves, natural gamma curves, and spontaneous potential curves. Based on existing technology and research, these four curves can respectively reflect the electrical conductivity, radioactivity, and electrochemical properties of rocks, and are highly correlated with lithology (sandstone, mudstone, etc.) and porosity, serving as the core basis for sedimentary facies identification. Therefore, if the processed well logging data contains these four curves, it will definitely undergo the aforementioned two normalization processes to serve as samples for subsequent multi-input model learning. In actual work, operators can, according to work needs, normalize other curve samples in the well logging data together with the above four curves to use as samples for subsequent multi-input model training.
[0026] Step S3: Create multiple data samples X1, X2...Xn based on the data processed in Step S2, and construct a dataset {X1i, X2i, ..., Xni} based on the established samples. At least one sample in the dataset must be a sandstone parameter sample. The staff will divide the processed data according to the parameter type. After division, parameters of the corresponding type will be selected as samples for training the multi-input model based on work needs and the work plan, thus determining the number of data samples to be created. When constructing the dataset, it is important to note that each data point {X1i, X2i, ..., Xni} within the dataset represents the data from the i-th well and corresponds to the sedimentary facies label of the i-th well. When the data {X1i, X2i, ..., Xni} is input into the multi-input model for training, the multi-input model will learn the correspondence between the input data {X1i, X2i, ..., Xni} and its corresponding sedimentary facies label. A comparison of existing methods for determining sedimentary facies using deep learning models with the sedimentary facies identification processes of experts in related fields reveals that experts consider not only the well logging data used during model training but also sandstone parameters as one of the criteria for sedimentary facies identification. Furthermore, existing research indicates a strong correlation between sandstone parameters and sedimentary facies. Therefore, including sandstone parameters as one of the data samples allows the learning model to incorporate them as a criterion for sedimentary facies identification, thereby improving the accuracy of sedimentary facies identification. Sandstone parameters typically include four data points: sandstone thickness, effective thickness, porosity, and permeability. Researchers can adjust the types of data within these parameters according to their specific needs and operational requirements.
[0027] In step S3, the data samples include well logging curves for the X1 sandstone section, X2 sandstone parameters, and X3 formation section. Generally, for sedimentary facies identification in ordinary areas, only three data samples (X1 sandstone section well logging curve, X2 sandstone parameters, and X3 formation section well logging curve) need to be established, and a dataset needs to be created to train the multi-input model to meet the conditions for sedimentary facies identification. For areas with special stratigraphic environments or special working environments, operators can adjust the number and types of data samples according to work needs, thereby adapting to sedimentary facies identification work under different environments.
[0028] It is important to note that after the samples are constructed, they should be classified according to sedimentary facies and the number of samples in each stratum should be counted. After classifying the samples according to sedimentary facies, if the number of samples corresponding to a certain sedimentary facies is small, the accuracy of the judgment results for that sedimentary facies may be poor during training due to insufficient training samples. Therefore, it is necessary to pay attention to the sample results that are judged as belonging to that sedimentary facies in subsequent training and validation processes.
[0029] Step S4: Construct a multi-input model. The multi-input model has multiple input branches, the number of which corresponds to the number of samples. Each input branch selects a sub-learning model corresponding to the input data. After inputting the data, the outputs of each input are combined to output the classification result. When building a multi-input model, it is important to note that each input branch corresponds to one data sample. For example, if the dataset obtained in step S3 is {X1i, X2i, ..., Xni}, the corresponding multi-input model should have n input branches. When inputting the dataset into the multi-input model for training, data X1i is input into the first input branch, data X2i into the second input branch, and so on until data Xni is input into the nth input branch. Since some data in the dataset is input in text form and some data is input in image form, and different training models have different learning effects for different types of data such as numbers and images, in order to improve the final accuracy of the sub-learning model, each input branch needs to select a model with better learning effect according to the type of input data it corresponds to. After the dataset is input to the input end of the multi-input model, each input branch will learn and train its input model separately and output a judgment result. At this time, the output results of each input branch are merged at the input end of the multi-input model to form a complete judgment result.
[0030] The sub-learning models include CNN and RNN models. CNN models extract local features (such as curve inflection points and amplitude differences) through convolutional layers and reduce dimensionality through pooling layers, making them suitable for training image data. RNN models are suitable for processing sequence data, such as time series data and natural language text. RNNs can utilize their recurrent structure to process this sequence information and better understand the dynamic changes of curves. During the process, each input branch selects an appropriate learning model based on the type of input data.
[0031] It is important to note that in multi-input models, the outputs of each sub-learning model at the input end are merged using a concatenation technique before the final calculation result is output. Generally, the sub-learning models within each input branch output results as vectors after receiving input data. The concatenation technique is used to concatenate these vectors into a new vector, which is then input into a classifier for determining the depositional phase.
[0032] Step S5: Input the dataset obtained in Step S3 into the multi-input model constructed in Step S4 for training. Divide the sample dataset obtained in Step S3 into a training set, a validation set, and a test set in a 6:2:2 ratio. During training, substitute the contents of the training set into the multi-input model for training and learning. Use the validation set to verify and correct the learning results of the multi-input model. During training, each data in the training set is processed by the input branch, and the corresponding parameter features are output by each input branch. For example, if {X11, X21, ..., Xn1} is input into the multi-input model, the output results of the input branches are feature 1, feature 2, feature 3... feature n, respectively. The output results of the input branches are concatenated to form {feature 1, feature 2, ..., feature n} using a concatenation technique. Assuming the corresponding sedimentary facies label is channel facies, the classifier connects {feature 1, feature 2, ..., feature n} with the channel facies. Repeat the above process until the training set is completed, and then input the data of the validation set into the trained multi-input model for verification. After the validation set data is input into the multi-input model, assuming there are three sedimentary facies: x, y, and z, the classifier will determine the probability {a, b, c} of the sedimentary facies corresponding to the validation set input data as x, y, or z based on the feature values contained in the received concatenated vector. Here, a, b, and c represent percentages of probability, and a + b + c = 1. The classifier selects the sedimentary facies with the highest probability among the three as the output result. When the sedimentary facies label corresponding to the validation set data is inconsistent with the output result, for example, the classifier outputs x but the actual sedimentary facies is y, the classifier will learn from this result. In subsequent classification processes, it will increase the probability of sedimentary facies of class y and decrease the probability of sedimentary facies of class x in similar feature combinations.
[0033] The classifier is actually a new training model established separately as the backbone network of the model in this experiment, which is formed by concatenating the vector results output by the output end. Similar to the selection of sub-training models in the input branch, the multi-input model selects an RNN or CNN model as the backbone network to learn the concatenated vector.
[0034] In addition, to ensure the accuracy of the final model, RNN and CNN models can be trained separately as backbone networks, and the model with higher accuracy can be selected to obtain the final multi-input model.
[0035] Step S6: Evaluate the model's training results. Input the test set into the trained model and compare the model's output with the actual results. Judge the model's training effect based on the accuracy between the two. Generally, input the data in the test set into the learning model and compare the model's output with the corresponding sedimentary phase labels in the test set. Generally, when the accuracy reaches 90% or higher, it indicates that the model training is complete and the training effect is good.
[0036] Step S7: Input the data to be predicted into the model for predictive analysis. After the model training is complete, when it is necessary to determine the sedimentary facies in the target area, simply input the corresponding data into the trained model, and the model will automatically output the classification results of the sedimentary facies.
[0037] Example 2: Figure 2-6 As shown, taking an oilfield X area as an example, the specific implementation steps of the present invention are explained in detail with reference to Embodiment 1.
[0038] Step S1: Collect formation and production well data within the target area; after organizing the collected test data, a total of 789 wells within the target area were collected, with a logging depth of approximately 300m and a logging sampling interval of 0.05m. Simultaneously, the sandstone data table was expanded based on the collected data, as follows: Figure 4 As shown. At this time, the facies field in the sandstone data table is the sedimentary facies field. The classification results of the collected data based on the data in this field are: channel facies H: 2511 samples, sandbar SB: 639 samples, interchannel sands S: 3008 samples.
[0039] Step S2: Analyze and clean the collected data to eliminate outliers; the distribution of the processed data is as follows. Figure 3 As shown.
[0040] Step S3: Create multiple data samples X1, X2...Xn based on the data processed in Step S2, and construct a dataset {X1i, X2i, ..., Xni} based on the established samples. At least one sample in the data samples is a sandstone parameter sample. In this embodiment, three data samples are established: sandstone section logging curve sample X1, sandstone section parameter sample X2, and formation section logging curve sample X3. Specifically, sandstone section logging curve sample X1 and formation section logging curve sample X3 include corresponding micro-gradient logging curves, micro-potential logging curves, natural gamma logging curves, and spontaneous potential logging curves, respectively. Sandstone section parameter sample X2 includes sandstone thickness, effective thickness, porosity, and permeability. The curves in sandstone section logging curve sample X1 and formation section logging curve sample X3 are converted as follows... Figure 4 As shown.
[0041] Step S4: Construct a multi-input model. The multi-input model has multiple input branches, the number of which corresponds to the number of samples. Each input branch selects a sub-learning model corresponding to the input data. After inputting the data, the outputs of each input are combined to output the classification result. During the learning process, the adaptive moment estimator (Adam) optimizer is used in training, with a learning rate of 0.0001, a batch size of 32, and a training batch size of 60 epochs. The constructed model architecture diagram is shown below. Figure 5-6 As shown, in this embodiment, in order to compare the two models, CNN and RNN, as the model skeleton, the BiLSTM model is selected for the RNN model and the ResNet50 model is selected for the CNN model. In the subsequent process, the two models can be trained separately, and the model with higher accuracy can be selected to determine the depositional phase of the target region based on the training effect.
[0042] Step S5: Input the dataset obtained in step S3 into the multi-input model constructed in step S4 for training; Step S6: Evaluate the training results of the multi-input model; After evaluation, the accuracy of the learning models using CNN and RNN as backbones were 92.85% and 92.60% respectively. The difference between the two is not large, and both can be used as the final models for sedimentary phase determination. This also shows that the improvement in accuracy mainly depends on the establishment of the multi-input model, rather than the selection of model type.
[0043] Step S7: Input the data to be predicted into the multi-input model for predictive analysis.
[0044] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A phase determination method based on well logging curves of sandstone and formation sections, characterized in that, include: Step S1: Collect formation and production well data within the target area; Step S2: Analyze and clean the data collected in step S1 to eliminate outliers; Step S3: Create multiple data samples X1, X2...Xn based on the data processed in step S2, and construct a dataset {X1i, X2i, ..., Xni} based on the established samples. At least one sample in the data samples is a sandstone parameter sample. Step S4: Construct a multi-input model. The multi-input model has multiple input branches at the input end. The number of input branches corresponds to the number of samples. Within each input branch, select the sub-learning model corresponding to the input data, and merge the output results of each input end to output the classification result. Step S5: Input the dataset obtained in step S3 into the multi-input model constructed in step S4 for training; Step S6: Evaluate the training results of the multi-input model; Step S7: Input the data requiring phase determination into the trained multi-input model for predictive analysis.
2. The phase determination method based on well logging curves of sandstone and formation sections according to claim 1, characterized in that, In step S1, the sandstone data table in the target area is expanded based on the collected data, and the fields of bottom depth and thickness of the strata are added.
3. The phase determination method based on well logging curves of sandstone and formation sections according to claim 1, characterized in that, In step S2, data with invalid and missing values are cleaned up, and then the remaining data is normalized.
4. The phase determination method based on well logging curves of sandstone and formation sections according to claim 3, characterized in that, The normalization process scales the logging curves to a range of {0-255} pixels.
5. The phase determination method based on well logging curves of sandstone and formation sections according to claim 4, characterized in that, The logging curves include microgradient curves, micropotential curves, natural gamma curves, and natural potential curves.
6. The phase determination method based on well logging curves of sandstone and formation sections according to claim 1, characterized in that, The data samples in step S3 include the X1 sandstone section logging curve, the X2 sandstone parameters, and the X3 formation section logging curve.
7. The phase determination method based on well logging curves of sandstone and formation sections according to claim 5, characterized in that, After the samples are constructed in step S3, the obtained samples are classified according to sedimentary facies and the number of samples of each type of stratum is counted.
8. The phase determination method based on well logging curves of sandstone and formation sections according to claim 3, characterized in that, In step S4, the sub-learning models include CNN models and RNN models. After the input of the multi-input model, the output results of each learning model at the input end are merged by the splicing technique to output the calculation result.
9. The phase determination method based on well logging curves of sandstone and formation sections according to claim 1, characterized in that, In step S5, the sample dataset obtained in step S3 is divided into a training set, a validation set, and a test set in a ratio of 6:2:
2. During the training process, the content of the training set is substituted into the multi-input model for training and learning. The learning results of the multi-input model are verified and corrected through the validation set.
10. The phase determination method based on well logging curves of sandstone and formation sections according to claim 9, characterized in that, In step S6, the test set is input into the trained multi-input model, and the output of the multi-input model is compared with the actual result. The training effect of the multi-input model is judged based on the accuracy between the two.
