Multi-view well lithology identification method for incomplete data set

Abstract

The invention provides a multi-view well lithology identification method for an incomplete data set, and relates to the technical field of logging lithology automatic identification, and the method comprises the steps: respectively selecting a proper preprocessing method, and dividing a view into an autocorrelation view and a non-correlation view according to the data feature correlation of different views, calculating a Manhattan distance between the to-be-classified sample of each view and all sample points to obtain a distance vector; for non-relevant views, deleting samples with missing data, and calculating a distance after view missing is caused; normalizing the distance vector of each view, and calculating according to a PDS thought to obtain a collaborative distance; selecting K samples with minimum collaborative distances with the to-be-classified samples as neighbor subsets; and finally, outputting the category with the highest occurrence frequency in the neighbor subsets as a classification result of the to-be-classified samples. According to the method, the mining degree of the lithology information in the logging information can be effectively improved, and the lithology identification effect is further improved.

Description

Technical Field

[0001] The invention relates to the technical field of automatic identification of well logging lithology, and in particular to a multi-view well lithology identification method for incomplete data sets. Background Art

[0002] Using various physical property data in the well to classify lithology is an important part of oil exploration. It can not only provide rich information for geological research, but also improve the effect of reservoir evaluation. In recent years, the application of machine learning in exploration, development, production, reservoir engineering and management planning in the oil industry has increased due to its advantages of automatic discrimination and rapid decision-making, and its application in lithology identification has gradually matured. However, with the advent of the big data era and the widespread attention paid to intelligent interpretation methods, existing lithology identification methods have begun to show their limitations.

[0003] Due to different ages or geological purposes, different logging series are measured, resulting in inconsistent numbers of logging curves in different wells, or due to various objective factors such as instrument instability and wellbore collapse, the logging curves of some well sections are distorted. These problems lead to incomplete logging data sets used for lithology identification, and most machine learning methods cannot directly process incomplete data sets. Although decision tree-based machine learning algorithms, such as XGBoost and LightGBM, can directly use incomplete data sets for model training and application, in the application stage, the algorithm will automatically divide missing values ​​into default nodes, making the results unreliable.

[0004] At present, people usually have three ways to deal with lithology identification of incomplete data sets: (1) Directly delete incomplete samples or features in the data set, but a large amount of lithological information is lost.

[0005] (2) Use samples of the same type as reference to fill in missing values. However, this method cannot be used when all samples of a certain type of lithology are missing a certain characteristic value (logging curve).

[0006] (3) Curve prediction technology is used to indirectly fill in the missing values ​​in the logging data set. However, the complexity of the model constructed by this technology far exceeds the lithology identification model itself, which greatly increases the cost.

[0007] On the other hand, logging data are in various forms, with different sources, dimensions, and even modes. These logging data describe lithology characteristics from different angles, and complement each other in the process of lithology identification to enhance the accuracy and reliability of lithology identification. Such data is called multi-view data in the field of machine learning. However, the existing lithology identification methods are mainly based on single-view methods, which are difficult to effectively integrate multi-view data. Based on this, the existing lithology identification methods are difficult to fully utilize logging data for lithology identification, and the utilization rate of information is low. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a multi-view well lithology identification method for incomplete data sets, which effectively improves the degree of mining lithology information in well logging data, thereby improving the lithology identification effect.

[0009] In order to solve the above technical problems, the technical solution of the present invention is as follows: In a first aspect, a multi-view well lithology identification method for an incomplete data set comprises: Acquire a well logging data set, and perform preprocessing to obtain preprocessed well logging data; According to the characteristics of well logging data, the well logging data is divided into different view data; According to the correlation of data features of different views, the views are divided into autocorrelated views and non-correlated views. The Manhattan distance between the sample to be classified and all sample points of each view is calculated to obtain the distance vector. For the autocorrelated view, the distance is calculated based on the partial distance strategy. For the non-correlated view, the samples with missing data are deleted, and the distance is calculated after the view is missing. Normalize the distance vector of each view separately, and calculate the collaborative distance according to the PDS idea; Select the samples with the smallest collaborative distance with the sample to be classified as the nearest neighbor subset; finally, output the category with the highest frequency in the nearest neighbor subset as the classification result of the sample to be classified.

[0010] Further, a well logging data set is obtained and preprocessed to obtain preprocessed well logging data, including: Perform preliminary data cleaning on the acquired logging data set and analyze the missing values ​​in the data set to obtain the results of missing value analysis; Based on the results of missing value analysis, the distribution of each category in the data set was analyzed to identify the lithology categories with small numbers; For each minority class, two samples in the same cluster are randomly selected; between the two selected minority class samples, artificial synthetic samples are generated in the feature space by linear interpolation, and the above process is repeated until the number of lithology samples of all categories is balanced; The synthesized new samples are added to the original dataset to eventually form a preprocessed complete dataset.

[0011] Furthermore, the views are divided into autocorrelated views and non-correlated views according to the correlation of data features of different views, and the Manhattan distance between the sample to be classified in each view and all sample points is calculated to obtain a distance vector, including: According to the correlation between the data features of different views, the views are divided into autocorrelation views and non-correlation views; the autocorrelation views contain highly correlated features, while the non-correlation views contain relatively independent features; For non-correlated views, if there are samples with missing data, these samples are deleted, resulting in missing views; In the non-correlated view, the Manhattan distance between the sample to be classified and all sample points is calculated by the following formula: ; in, represents Manhattan distance; and Represents the eigenvectors With the eigenvector No. eigenvalues; For calculation The number of features used.

[0012] Furthermore, the distance vector of each view is normalized respectively, and the collaborative distance is calculated according to the PDS idea, including: For each view, the calculated Manhattan distance vector is normalized to eliminate the distance metric differences between different views due to different feature scales; After normalizing the distance vector of each view, the collaborative distance is calculated according to the PDS idea. The specific calculation formula is: ; in, and Represents the eigenvectors With the eigenvector No. eigenvalues; For calculation The number of features used; is the total number of features including missing features; is the coordination distance.

[0013] Furthermore, between the two selected minority class samples, artificial synthetic samples are generated in the feature space by linear interpolation, including: Randomly select two samples from the minority class samples. The two samples should be in the same cluster, that is, belong to the same lithology category, so that the relative position of the selected samples in the feature space can represent the data distribution of this category; The feature space of the two samples is composed of multiple dimensions, each dimension corresponds to a specific lithological feature; linear interpolation is performed on the corresponding dimension to generate new points in the feature space, wherein the linear interpolation process specifically includes: For each pair of selected samples, a point is randomly selected on the line connecting them as the location of the synthetic sample; the random point is obtained by interpolating between the two samples with a random number between 0 and 1 as the proportional factor, and the interpolation result represents a weighted combination of the features of the two original samples; According to the interpolation result, a new synthetic sample is created, which contains all the eigenvalues ​​calculated by interpolation, thereby forming a new data point in the feature space; The iteration is repeated according to the number of synthetic samples to be generated, and a new synthetic sample is generated in each iteration; after all iterations are completed, the generated synthetic samples are added to the original dataset.

[0014] Furthermore, lithological characteristics include component contents and physical properties.

[0015] Furthermore, the samples with the smallest collaborative distance with the sample to be classified are selected as the nearest neighbor subset; finally, the category with the highest frequency in the nearest neighbor subset is output as the classification result of the sample to be classified, including: Sort the calculated collaborative distances from small to large to find the sample closest to the sample to be classified; According to the preset number of nearest neighbors K, select the K samples with the closest distance from the sorted distance list, and the samples constitute the nearest neighbor subset of the sample to be classified; For the selected neighbor subset, traverse each sample in it and record their respective category labels. Use a counter to count the frequency of each category appearing in the neighbor subset. After counting the class frequencies of all neighboring samples, find the class with the highest frequency; if there are multiple classes with the same and highest frequency, select one of them according to specific needs; The category with the highest frequency of occurrence is determined as the predicted category of the sample to be classified, and the result is output.

[0016] The above solution of the present invention includes at least the following beneficial effects: The method proposed in this patent can directly process incomplete lithology identification data sets, effectively improving the utilization rate of data. Different views in this patent method can take different preprocessing operations for different views according to needs; as a multi-view method, this patent method can simultaneously use logging data from multiple different sources and different modes for lithology identification; the improved view fusion strategy of this patent method can more effectively fuse multi-view data; this patent method has good robustness to missing curves in the well; this patent method can improve the lithology identification effect of complex lithology reservoirs to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1It is a flowchart of a multi-view well lithology identification method for an incomplete data set, namely, PDS-MVKNN, provided by an embodiment of the present invention.

[0018] Figure 2 It is a porosity view feature correlation analysis diagram of an embodiment of the present invention.

[0019] Figure 3 It is a characteristic correlation analysis diagram of a radioactive view according to an embodiment of the present invention.

[0020] Figure 4 This is a simplified SMOTE method sample expansion effect diagram of the present invention (taking andesite as an example).

[0021] Figure 5 It is a line graph of different model accuracy rates under different K values ​​in an embodiment of the present invention.

[0022] Figure 6 It is a confusion matrix diagram of model a in the embodiment of the present invention.

[0023] Figure 7 It is a confusion matrix diagram of model b in the embodiment of the present invention.

[0024] Figure 8 It is a comparison of the accuracy of different methods under different missing rates of the present invention.

[0025] Fig. 9 It is a comparison diagram of lithology identification results of Well X-1 simulating the missing curve situation in different wells in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0027] An embodiment of the present invention provides a multi-view well lithology identification method for an incomplete data set, the method comprising the following steps: Acquire a well logging data set, and perform preprocessing to obtain preprocessed well logging data; According to the characteristics of well logging data, the well logging data is divided into different view data; According to the correlation of data features of different views, the views are divided into autocorrelated views and non-correlated views. The Manhattan distance between the sample to be classified and all sample points of each view is calculated to obtain the distance vector. For the autocorrelated view, the distance is calculated based on the partial distance strategy. For the non-correlated view, the samples with missing data are deleted, and the distance is calculated after the view is missing. Normalize the distance vector of each view separately, and calculate the collaborative distance according to the PDS idea; Select the samples with the smallest collaborative distance with the sample to be classified as the nearest neighbor subset; finally, output the category with the highest frequency in the nearest neighbor subset as the classification result of the sample to be classified.

[0028] In an embodiment of the present invention, by reasonably dividing autocorrelation views and non-correlation views, the method can more effectively utilize incomplete data in logging data. For autocorrelation views, the partial distance strategy (PDS) is used to calculate the distance, so that useful information can be retained even in the case of missing data. For non-correlation views, the interference of incomplete data on distance calculation is avoided by deleting samples with missing data and then calculating the distance, thereby improving the utilization rate of data. The method calculates the Manhattan distance of each view and obtains the collaborative distance by normalization, which can more accurately reflect the similarity between the sample to be classified and the known sample. The sample with the smallest collaborative distance with the sample to be classified is selected as the nearest neighbor subset, which further improves the accuracy of classification. At the same time, the category with the highest frequency in the nearest neighbor subset is output as the classification result, which helps to reduce the influence of accidental errors and makes the classification result more stable and reliable. The method is suitable for processing incomplete data sets and has good adaptability to the problem of missing data that often occurs in actual logging processes. By reasonably processing incomplete data, the method can reduce the requirements for data integrity while ensuring classification accuracy, thereby improving the practical application value of the method. By optimizing the distance calculation and the nearest neighbor subset selection process, this method can improve the computational efficiency while ensuring the classification effect. This is of great significance for processing large-scale logging data sets, helping to shorten the data processing cycle and improve work efficiency.

[0029] Further, a well logging data set is obtained and preprocessed to obtain preprocessed well logging data, including: Perform preliminary data cleaning on the acquired logging data set and analyze the missing values ​​in the data set to obtain the results of missing value analysis; Fill missing values ​​based on the results of missing value analysis; analyze the distribution of various categories in the data set and identify lithology categories with small numbers; For each minority class, two samples in the same cluster are randomly selected; between the two selected minority class samples, artificial synthetic samples are generated in the feature space by linear interpolation, and the above process is repeated until the number of samples in all categories is balanced; The synthesized new samples are added to the original dataset to eventually form a preprocessed complete dataset.

[0030] In an embodiment of the present invention, through preliminary data cleaning, repetition, abnormality and non-critical information in the data set can be removed, thereby improving the accuracy and consistency of the data. This helps subsequent data analysis and model building, making the results more reliable. In-depth analysis of the missing values ​​in the data set and appropriate filling according to the analysis results can effectively solve the problems caused by the missing values. This can not only reduce the loss of data information, but also ensure the integrity of the data set, and provide convenience for subsequent data processing. By analyzing the distribution of each category in the data set and identifying the lithology categories with a small number, samples of these minority categories can be synthesized in a targeted manner. This helps to improve the category balance of the data set, improve the model's recognition ability for each category, and avoid the model from being overly biased towards the majority class during the training process. By generating artificial synthetic samples in the feature space by linear interpolation, the diversity of the data set can be increased. These synthetic samples can simulate situations that may occur in practice but are not covered in the original data set, thereby improving the generalization ability of the model. Adding the synthesized new samples to the original data set can not only expand the size of the data set, but also optimize the distribution of the data set. This helps the model learn more features and patterns during the training process and improves the performance and stability of the model.

[0031] Furthermore, the views are divided into autocorrelated views and non-correlated views according to the correlation of data features of different views, and the Manhattan distance between the sample to be classified in each view and all sample points is calculated to obtain a distance vector, including: According to the correlation between the data features of different views, the views are divided into autocorrelation views and non-correlation views; the autocorrelation views contain highly correlated features, while the non-correlation views contain relatively independent features; For non-correlated views, if there are samples with missing data, these samples are deleted, resulting in missing views; In the non-correlated view, the Manhattan distance between the sample to be classified and all sample points is calculated by the following formula: ; in, represents Manhattan distance; and Represents the eigenvectors With the eigenvector No. eigenvalues; For calculation The number of features used.

[0032] In an embodiment of the present invention, by dividing the views into two categories, autocorrelation and non-correlation, different features can be processed more specifically. For highly correlated features (autocorrelation views), the intrinsic connection between them can be considered, while for relatively independent features (non-correlation views), their influence can be considered separately. This helps to more accurately characterize data features and improve the accuracy of classification. In non-correlation views, if there are samples with missing data, by deleting these samples, calculation errors caused by missing data can be avoided. This approach can ensure that the data used when calculating the Manhattan distance is complete and reliable, thereby improving the accuracy of distance calculation. Manhattan distance, as a commonly used distance measurement method, has good robustness when calculating the similarity between samples to be classified and known samples. It takes into account the differences in feature vectors in various dimensions and can more comprehensively reflect the similarity between samples. By calculating the Manhattan distance, the distance vector between the sample to be classified and all sample points in each view can be obtained. These distance vectors can be used as important inputs of the classifier to help the classifier more accurately determine the category of the sample to be classified. At the same time, since the Manhattan distance calculation is relatively simple and efficient, the efficiency of classification can be improved to a certain extent.

[0033] Furthermore, the distance vector of each view is normalized respectively, and the collaborative distance is calculated according to the PDS idea, including: For each view, the calculated Manhattan distance vector is normalized to eliminate the distance metric differences between different views due to different feature scales; After normalizing the distance vector of each view, the collaborative distance is calculated according to the PDS idea. The specific calculation formula is: ; in, and Represents the eigenvectors With the eigenvector No. eigenvalues; For calculation The number of features used; is the total number of features including missing features; is the coordination distance.

[0034] In an embodiment of the present invention, by normalizing the Manhattan distance vector calculated for each view, the distance metric differences between different views caused by different feature scales can be effectively eliminated. This ensures that different views have the same weight when calculating the collaborative distance, thereby improving the accuracy and reliability of the distance metric. The calculation of the collaborative distance integrates information from multiple views and fully considers the complementarity and consensus between different views. This helps to more comprehensively characterize the relationship between the samples to be classified and the known samples, and improves the accuracy and robustness of the classification decision. The PDS idea allows the influence of missing features to be considered when calculating the collaborative distance. By introducing the total feature quantity m including the missing features, the calculation of the collaborative distance can more truly reflect the actual differences between the data, and maintain good performance even in the presence of missing data. The normalization and collaborative distance calculation methods help reduce the model's dependence on a specific data set and improve the generalization ability of the model. This means that the model can still maintain a good classification effect when processing data sets with different distributions or feature scales.

[0035] Furthermore, between the two selected minority class samples, artificial synthetic samples are generated in the feature space by linear interpolation, including: Randomly select two samples from the minority class samples. The two samples should be in the same cluster, that is, belong to the same lithology category, so that the relative position of the selected samples in the feature space can represent the data distribution of this category; The feature space of the two samples is composed of multiple dimensions, each dimension corresponds to a specific lithological feature; linear interpolation is performed on the corresponding dimension to generate new points in the feature space, wherein the linear interpolation process specifically includes: For each pair of selected samples, a point is randomly selected on the line connecting them as the location of the synthetic sample; the random point is obtained by interpolating between the two samples with a random number between 0 and 1 as the proportional factor, and the interpolation result represents a weighted combination of the features of the two original samples; According to the interpolation result, a new synthetic sample is created, which contains all the eigenvalues ​​calculated by interpolation, thereby forming a new data point in the feature space; The iteration is repeated according to the number of synthetic samples to be generated, and in each iteration, a new synthetic sample is generated; after all iterations are completed, the generated synthetic samples are added to the original data set; the lithological characteristics include component content and physical properties.

[0036] In an embodiment of the present invention, in a data set, the number of samples of certain lithology categories may be far less than that of other categories, resulting in insufficient learning of these minority categories by the model during training. By generating artificial synthetic samples through linear interpolation, the number of samples of minority classes can be effectively increased, thereby improving the class balance of the data set and enabling the model to learn the characteristics of each category more comprehensively. Linear interpolation creates new data points in the feature space, which represent weighted combinations of the original sample features. This approach can increase the diversity of the feature space, allowing the model to be exposed to a wider range of feature combinations during training, thereby improving the generalization ability of the model. Since linear interpolation is performed between two selected minority class samples, and the two samples are in the same cluster, the generated synthetic samples can better maintain the continuity of the original data distribution. This helps the model to more accurately capture the boundaries and transition areas between lithology categories. By increasing the number of samples of minority classes and the diversity of feature space, the model can more fully learn the characteristics of each category during training, thereby reducing the bias towards the majority class. This will help improve the performance of the model in identifying minority classes and improve the overall classification accuracy. During the interpolation process, multiple characteristics of lithology are considered, including component content and physical properties. This ensures that the generated synthetic samples are consistent with the original samples in terms of characteristics, allowing these characteristics to be more effectively used for lithology identification.

[0037] Furthermore, the samples with the smallest collaborative distance with the sample to be classified are selected as the nearest neighbor subset; finally, the category with the highest frequency in the nearest neighbor subset is output as the classification result of the sample to be classified, including: Sort the calculated collaborative distances from small to large to find the sample closest to the sample to be classified; According to the preset number of nearest neighbors K, select the K samples with the closest distance from the sorted distance list, and the samples constitute the nearest neighbor subset of the sample to be classified; For the selected neighbor subset, traverse each sample in it and record their respective category labels. Use a counter to count the frequency of each category appearing in the neighbor subset. After counting the class frequencies of all neighboring samples, find the class with the highest frequency; if there are multiple classes with the same and highest frequency, select one of them according to specific needs; The category with the highest frequency of occurrence is determined as the predicted category of the sample to be classified, and the result is output.

[0038] In an embodiment of the present invention, by selecting K samples with the smallest collaborative distance as the nearest neighbor subset, the number of samples to be considered can be greatly reduced, thereby simplifying the computational complexity in the classification process. This neighbor-based method is computationally efficient and is particularly suitable for large-scale data sets. The method focuses on considering the local area around the sample to be classified, that is, its nearest K neighbors. This local consideration helps to capture the local structure and pattern of the data, and is particularly effective for complex or nonlinearly separable data sets. By counting the most frequently occurring categories in the nearest neighbor subset to determine the predicted category of the sample to be classified, this method can reduce the impact of noise data or outliers on the classification results to a certain extent. Because even if there are individual erroneous neighbor samples, their categories are unlikely to become the most frequently occurring categories. The classification method based on the nearest neighbor is intuitive and easy to understand and easy to explain. Outputting the category with the highest frequency in the nearest neighbor subset as the result, this decision-making method conforms to human intuitive cognition and helps to enhance the credibility and acceptability of the model. The method does not require strict assumptions or modeling of the data set in advance, but determines the nearest neighbor subset and the predicted category based on the actual distribution of the data. This adaptability enables the method to flexibly handle data sets of various types and distributions.

[0039] Take the lithology identification of complex volcanic reservoirs in a certain area as an example. The Mesozoic ancient buried hill igneous reservoirs in the region are developed in large quantities in the Pearl River Mouth Basin. They were affected by multiple periods of magmatic activity during their formation. They are distributed from basic to neutral to acidic, and weathering and alteration phenomena are common. The lithology is complex and diverse. Some lithology logging response characteristics are similar, and it is difficult to accurately identify them using only logging data. Since the minerals in specific strata in the same area are composed of relatively fixed chemical elements, the strata in different regions and different layers have different lithology characteristics and mineral compositions, and the corresponding element types and contents also have certain differences. Therefore, the element indication effect on lithology is used to add the cuttings logging element data to the data set to improve the lithology identification effect.

[0040] Based on thin section identification, a total of 263 identifiable lithological samples of 10 different logging scales were collected from 16 wells, including andesite, altered andesite, granite, two different weathered granites, tectonic schist, diorite, altered diabase, granodiorite and altered granite.

[0041] This embodiment does not analyze the importance of features, and directly selects as many lithology sensitivity curves as possible. In this area, this paper selects three curves reflecting the porosity of the formation, namely neutron (CN), density (DEN), and Delta T (DT) as the porosity view, five curves reflecting the radioactivity of the formation, namely natural gamma ray (GR), uranium (U), thorium (Th), potassium (K40), and gamma ray without uranium (KTh), as the radioactivity view, and the element logging curve reflecting the composition of the formation as the element view. Figure 2 and Figure 3 This is a correlation analysis diagram of logging response characteristics. According to the scatter plot and density plot and combined with the measurement principles of each curve, it can be known that the porosity view and radioactivity view are both autocorrelation views, among which the density feature in the porosity view needs to be negative before use.

[0042] Table 1 is a brief table of the data set. It can be seen that the lithological data of andesite and granite are seriously missing, and the element views of the two weathered granites are all missing. The comprehensive missing rate is 10.43%. In addition, the data volume is small and the number of samples of each type is seriously unbalanced.

[0043]

[0044] The three curves of neutron (CN), density (DEN) and acoustic wave (AC) reflecting the porosity of the formation are selected as the porosity view, the five curves of natural gamma (GR), uranium (U), thorium (TH), potassium (K40) and uranium-free gamma (KTH) reflecting the radioactivity of the formation are selected as the radioactivity view, and the element logging curve reflecting the formation composition is selected as the element view.

[0045] Taking the neutron-acoustic crossplot of andesite as an example, each category is expanded to 100 samples. Taking the neutron (CNL)-acoustic (DT) crossplot as an example, Figure 4 The simplified SMOTE method is used to balance the effect of different categories of data sets. The points on the nan axis are sample points with missing values. It can be seen that the simplified SMOTE method effectively increases the density of sample points. The comprehensive missing rate of the data after SMOTE is 16.69%.

[0046] There are dimensional differences between logging curves, and PDS-MVKNN, as a nearest neighbor algorithm, is very sensitive to feature scales, so it needs to be standardized. This paper uses the maximum and minimum normalization method: ; in To standardize the results; is the characteristic observation value; is the minimum value of the overall data of this feature; is the maximum value of the characteristic overall data. Since the DEN curve is negatively correlated with other porosity logging curves, this paper first calculates the negative of the DEN curve and then normalizes it.

[0047] Unlike other lithological characteristics, element logging technology can usually detect 14 major elements and 20 trace elements. Element logging measures a large number of elements but has low accuracy. Ten elements corresponding to core oxides in element logging data, including sodium (Na), magnesium (Mg), aluminum (Al), silicon (Si), phosphorus (P), potassium (K), calcium (Ca), titanium (Ti), manganese (Mn), and iron (Fe), are selected for standardization, which can reduce the error caused by different element types due to different element logging inversion models in different wells.

[0048] The formula for the relative content of these 10 elements after standardization is as follows: ; In the formula, is the relative percentage content of X element in mud logging, unit is %; is the logging data corresponding to the X element; It is the sum of the above 10 elements in the corresponding logging data.

[0049] Among the existing MVKNN methods, a method (Voting) that builds a classifier for each view separately and then votes on the classification results of each view to determine the final output can allow for missing views, while other methods cannot be used on incomplete data sets. Comparing the proposed method (PDS-MVKNN-PDS) with the improved Voting method (PDS-MVKNN) based on PDS can verify the effect of the improvement made by this patent on the view coupling method. The data set is subjected to stratified five-fold cross validation, and the accuracy, precision, recall and F1 values ​​are selected to evaluate the performance of the model. The accuracy reflects the ability of the model to correctly judge the overall sample, the precision reflects the accuracy of the positive samples in the prediction results, the recall reflects the ability of the model to correctly predict the full degree of positive samples, and the F1 value reflects the comprehensive performance of the model. The closer the four performance parameters are to 1, the better the model performance. Equations 6 to 9 are expressions of evaluation indicators.

[0050] ; ; ; ; Among them, TP represents the positive samples predicted by the model as positive, TN represents the negative samples predicted by the model as negative, FP represents the negative samples predicted by the model as positive, and FN represents the positive samples predicted by the model as negative.

[0051] Traverse K values ​​from 1-20 and keep the best model. Figure 5 The following is a line chart of the accuracy of different K values. As the K value increases, the PDS method is more affected by missing data and the recognition accuracy decreases. Therefore, the optimal K value is small, and the recognition accuracy is the highest when K=3. As the K value increases, the Voting method is less affected by missing data and the recognition accuracy increases. The accuracy is the highest when K=14.

[0052] Figure 6 is the confusion matrix of model a, Figure 7 For model b, the sample size of each category is 100 after oversampling. Therefore, it can be seen from the confusion matrix that the recognition accuracy of each category of the method proposed in this paper (model a) is as low as 80% and as high as 100%, and the recognition accuracy is relatively stable. Compared with the confusion matrix of PDS-MVKNN without using element view (model b), the addition of element view significantly enhances the performance of the lithology recognition model in this area and plays a good correction role for tectonic schist (lithology 6).

[0053] Table 2 shows the average performance of the five-fold cross validation of several models. The performance of this method is slightly higher than that of the Voting method.

[0054]

[0055] Figure 8 The accuracy of different methods under different deletion rates is compared. As the deletion rate increases, the recognition accuracy of the two classifiers decreases, but the decrease is slow. The recognition accuracy of the view fusion strategy proposed in this paper is higher than that of the integration method.

[0056] Well X-1 is a well with sufficient data and complex lithology in the study area. Fig. 9 The first channel is the depth, the second channel is the lithology identification result given by the present invention when all the features selected during modeling are used, and the sixth channel is the lithology determined manually based on the integrated thin section and logging information. The common missing curves in different regions are inconsistent. The acoustic wave and density curves are greatly affected by the borehole conditions and are often distorted. The radioactive spectrum logging is affected by the source usage requirements and has a high cost. Non-key wells such as production wells often abandon the radioactive spectrum logging series. In order to verify the tolerance of the present invention to the missing curves in the well during the model application stage, Fig. 9Tracks 3 to 5 respectively simulate the lithology identification results of the present invention for three common cases of missing well curves: (1) missing AC and DEN curves, (2) missing U, TH, K40, KTH four radioactive spectrum logging curves, and (3) missing all six curves. By comparison, it can be seen that missing well curves will reduce the identification effect. Compared with track 2, tracks 3 to 5 identify more erroneous interlayers. The fifth track has the most missing curves and the worst identification effect. The missing well curves reduce the lithology identification accuracy of PDSKNN to a certain extent, but still ensure the identification effect of most well sections, indicating that the present invention has good tolerance for missing well curves in the model application stage.

[0057] An embodiment of the present invention further provides a multi-view well lithology identification system for an incomplete data set, comprising: An acquisition module is used to acquire a well logging data set and perform preprocessing to obtain preprocessed well logging data; A preprocessing module is used to divide the data into different view data according to the characteristics of the preprocessed data; The partitioning module is used to partition the view and set the K value. For the autocorrelated view, PDS is used to calculate the distance between the sample to be classified and all sample points respectively. For the non-correlated view, the samples with missing data are deleted to cause the view missing, and the Manhattan distance between the sample to be classified and all sample points is calculated. The distance vector of each view is normalized respectively, and the collaborative distance is calculated according to the PDS idea. The K samples with the smallest distance to the sample to be classified are selected as the nearest neighbor subset; the corresponding category in the nearest neighbor subset is output.

[0058] It should be noted that the system is a system corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0059] The embodiment of the present invention further provides a computing device, comprising: a processor, a memory storing a computer program, wherein when the computer program is executed by the processor, the method described above is executed. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0060] The embodiment of the present invention also provides a computer-readable storage medium storing instructions, which, when executed on a computer, enable the computer to execute the method described above. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0061] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-view well lithology identification method for incomplete data sets, characterized in that: The method comprises: Acquire a well logging data set, and perform preprocessing to obtain preprocessed well logging data; According to the characteristics of well logging data, the well logging data is divided into different view data; According to the correlation of data features of different views, the views are divided into autocorrelated views and non-correlated views. The Manhattan distance between the sample to be classified and all sample points of each view is calculated to obtain the distance vector. For the autocorrelated view, the distance is calculated based on the partial distance strategy. For the non-correlated view, the samples with missing data are deleted, and the distance is calculated after the view is missing. Normalize the distance vector of each view separately, and calculate the collaborative distance according to the PDS idea; Select K samples with the smallest collaborative distance with the sample to be classified as the nearest neighbor subset; finally, output the category with the highest frequency in the nearest neighbor subset as the classification result of the sample to be classified.

2. The multi-view well lithology identification method for incomplete data sets according to claim 1, characterized in that: Acquire the well logging data set and perform preprocessing to obtain preprocessed well logging data, including: Perform preliminary data cleaning on the acquired logging data set and analyze the missing values ​​in the data set to obtain the results of missing value analysis; Based on the results of missing value analysis, the distribution of each category in the data set was analyzed to identify the lithology categories with small numbers; For each minority class, two samples in the same cluster are randomly selected; between the two selected minority class samples, artificial synthetic samples are generated in the feature space by linear interpolation, and the above process is repeated until the number of lithology samples of all categories is balanced; The synthesized new samples are added to the original dataset to eventually form a preprocessed dataset.

3. The multi-view well lithology identification method for incomplete data sets according to claim 2, characterized in that: According to the correlation of data features in different views, the views are divided into autocorrelated views and non-correlated views, and the Manhattan distance between the sample to be classified in each view and all sample points is calculated to obtain the distance vector, including: According to the correlation between the data features of different views, the views are divided into autocorrelation views and non-correlation views; the autocorrelation views contain highly correlated features, while the non-correlation views contain relatively independent features; For non-correlated views, if there are samples with missing data, these samples are deleted, resulting in missing views; In the non-correlated view, the Manhattan distance between the sample to be classified and all sample points is calculated by the following formula: ； in, represents Manhattan distance; and Represents the eigenvectors With the eigenvector No. eigenvalues; For calculation The number of features used.

4. The multi-view well lithology identification method for incomplete data sets according to claim 3, characterized in that: Normalize the distance vector of each view separately, and calculate the collaborative distance according to the PDS idea, including: For each view, the calculated Manhattan distance vector is normalized to eliminate the distance metric differences between different views due to different feature scales; After normalizing the distance vector of each view, the collaborative distance is calculated according to the PDS idea. The specific calculation formula is: ； in, and Represents the eigenvectors With the eigenvector No. eigenvalues; For calculation The number of features used; is the total number of features including missing features; is the coordination distance.

5. The multi-view well lithology identification method for incomplete data sets according to claim 4, characterized in that: Between the two selected minority class samples, artificial synthetic samples are generated in the feature space by linear interpolation, including: Randomly select two samples from the minority class samples. The two samples should be in the same cluster, that is, belong to the same lithology category, so that the relative position of the selected samples in the feature space can represent the data distribution of this category; The feature space of the two samples is composed of multiple dimensions, each dimension corresponds to a specific lithological feature; linear interpolation is performed on the corresponding dimension to generate new points in the feature space, wherein the linear interpolation process specifically includes: For each pair of selected samples, a point is randomly selected on the line connecting them as the location of the synthetic sample; the random point is obtained by interpolating between the two samples with a random number between 0 and 1 as the proportional factor, and the interpolation result represents a weighted combination of the features of the two original samples; According to the interpolation result, a new synthetic sample is created, which contains all the eigenvalues ​​calculated by interpolation, thereby forming a new data point in the feature space; The iteration is repeated according to the number of synthetic samples to be generated, and a new synthetic sample is generated in each iteration; after all iterations are completed, the generated synthetic samples are added to the original dataset.

6. The multi-view well lithology identification method for incomplete data sets according to claim 5, characterized in that: Lithological characteristics include component content and physical properties.

7. The multi-view well lithology identification method for incomplete data sets according to claim 6, characterized in that: Select the samples with the smallest collaborative distance with the sample to be classified as the nearest neighbor subset; Finally, the category with the highest frequency in the neighbor subset is output as the classification result of the sample to be classified, including: Sort the calculated collaborative distances from small to large to find the sample closest to the sample to be classified; According to the preset number of nearest neighbors K, select the K samples with the closest distance from the sorted distance list, and the samples constitute the nearest neighbor subset of the sample to be classified; For the selected neighbor subset, traverse each sample in it and record their respective category labels. Use a counter to count the frequency of each category appearing in the neighbor subset. After counting the class frequencies of all neighboring samples, find the class with the highest frequency; if there are multiple classes with the same and highest frequency, select one of them according to specific needs; The category with the highest frequency of occurrence is determined as the predicted category of the sample to be classified, and the result is output.

Images