Underground reservoir lithology identification method based on active learning technology

Abstract

The invention relates to an underground reservoir lithology identification method based on an active learning technology, and belongs to the field of data mining. The problem that the lithology of large-scale lithology data is difficult to identify due to lack of an efficient evaluation method is solved. According to the technical scheme, firstly, a neighbor diffusion network is generated, and the relation between lithology data nodes is organized; then, calculating topological structure importance and suspiciousness of lithology data nodes to obtain a key node sequence, performing lithology identification on a small part of lithology data nodes under the annotation of experts, and then annotating other lithology data nodes which are not annotated by the experts through a diffusion effect in combination with a neighborhood set; and a final lithology identification result is obtained after multiple rounds of iteration. According to the underground reservoir lithology identification method based on the active learning technology, lithology identification of large-scale lithology data is completed by labeling a small number of lithology data nodes.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas geological exploration, and in particular to an underground reservoir lithology identification method based on active learning technology. Background Art

[0002] With the continuous deepening of geological exploration, people need to understand the mineral composition and distribution of rocks to provide data support for geological research, mineral exploration and other work. How to process data is an extremely important issue. The traditional lithology identification method relies on direct observation and experience accumulation. Geologists analyze the various appearance characteristics of rocks through on-site surveys, and then manually divide the lithology based on stratigraphic principles. Although this method can effectively identify the lithology, it takes a lot of time and requires professionals in the industry to make judgments, which further limits the recognition efficiency. It is impossible to complete the lithology analysis of large-scale data by manpower alone. In the early twentieth century, the development of microscope technology brought new ideas to geologists. By observing rock thin sections to judge the mineral composition of rocks, the lithology can be identified more accurately. With the continuous maturity of drilling technology, people have obtained more accurate and comprehensive geological data, but there is still a lack of efficient ways to analyze them. It is still a technical problem to correctly process large amounts of data.

[0003] Entering the 21st century, the application of remote sensing technology, artificial intelligence and machine learning has provided new impetus for the development of lithology identification. Lithology identification has gradually developed towards digitalization and automation. Remote sensing image analysis technology helps people obtain a wide range of lithology information, while machine learning algorithms help process large amounts of data. The model is trained by taking part of the data as input, and then the trained model is used to give lithology identification results for the newly input lithology data. On this basis, people also combine observation slices and other methods for comprehensive analysis to obtain a more comprehensive and accurate lithology identification result. However, the existing technology still has some shortcomings: on the one hand, a large amount of labeled data is required for training the model, which is costly. Traditional machine learning methods require a large amount of lithology data to train the model, and have high requirements for resources and storage space. Especially when facing large-scale lithology data sets, a large amount of matrix operations and storage management may be required, without considering reducing resource consumption and improving efficiency while ensuring quality; on the other hand, the effect depends on the training data, which may lead to overfitting or underfitting problems. When training a complex model, it is very likely to learn the noise in the data set, resulting in good performance on the training data, but poor performance on the new data, which is overfitting. When the model is simple, it is difficult to capture the underlying patterns from the lithology data and thus cannot produce good results. Summary of the invention

[0004] In view of the shortcomings of the current technology mentioned in the background technology, the present invention proposes an underground reservoir lithology identification method based on active learning technology, the purpose of which is to combine active learning technology, improve the identification accuracy, save time and resources through a more effective lithology data processing method, and provide a new solution to the lithology identification problem in geological exploration. The present invention adopts the following technical solution to achieve the above purpose, including the following steps: S1. Collection and collation of original logging data and conventional logging curve data, and preprocessing of existing lithology data; S2, generating a neighbor diffusion network based on the preprocessed lithology data; S3. Dynamically update the criticality of lithology data nodes based on the neighbor diffusion network, and then obtain the critical node sequence And the query sequence ; S4. Based on the expert's judgment, the neighborhood set of lithology data nodes is updated, and a small number of lithology data nodes are identified and the relationship structure of the neighbor diffusion network is optimized; S5. Perform annotation diffusion based on expert constraints on the nearest neighbor diffusion network to obtain the lithology identification result of the current round; S6. Iteratively optimize the neighbor diffusion network relationship structure, stop iterating after reaching the specified number of iterations, and output the recognition result.

[0005] The beneficial effects of the present invention are as follows: the present invention generates a neighbor diffusion network and introduces an active learning framework; first, the original lithology data is preprocessed; second, the most critical lithology data is selected and handed over to experts for annotation; then, the annotations constrained by the experts are diffused on the basis of the neighbor diffusion network, thereby achieving the simultaneous conduct of model training and lithology identification; while saving time and improving the accuracy of lithology identification, the capital investment and cost are reduced, the efficiency is greatly improved, and a new method for lithology identification is provided.

[0006] Furthermore, the S1 comprises the following steps: S101. Obtain lithology data, including well diameter, gamma ray, natural potential, resistivity, acoustic wave, density, and photoelectric absorption index; S102, preprocessing the lithology data; filling in missing values ​​and extracting features.

[0007] The beneficial effects of the above further scheme are: the present invention improves the quality and reliability of the lithological data by preprocessing the acquired lithological data, and also enables the data to be better used; by extracting the features of more complex original lithological data, establishing standards for it, and quantifying its various features, while improving the data quality, it also provides a guarantee for the accuracy of subsequent analysis.

[0008] Further, the S2 comprises the following steps: S201, taking all the lithology data after preprocessing as nodes in a neighbor diffusion network; S202: Select a proxy node from the connected components formed by each lithology data node, and merge the connected components.

[0009] The beneficial effect of the above further scheme is that through a special composition method, the relationship between lithology data nodes is established from bottom to top, which saves space resources while ensuring the connectivity of the entire neighbor diffusion network; in addition, it also provides a medium for subsequent annotation diffusion based on expert constraints.

[0010] Furthermore, S3 comprises the following steps: S301. Calculate the topological importance of all nodes and generate representative sequences ; S302: Calculate the suspiciousness of all nodes and generate a suspiciousness sequence ; S303, based on representative sequences and suspicious degree sequence Generate key node sequence ; S304, according to the key node sequence Generate query sequence .

[0011] The beneficial effects of the above further scheme are: giving priority to some lithology data nodes with higher suspicion, while also taking into account some lithology data nodes with more critical topological structures. The advantage is that the information given by experts is efficiently used to better judge the more ambiguous and more difficult lithology data nodes. In addition, the accuracy of lithology identification is improved and resource consumption is reduced; by generating query sequences , the key lithology data are handed over to experts for labeling, which in turn guides the subsequent lithology identification process. While improving the accuracy of lithology identification, it also speeds up the convergence of the algorithm and improves the efficiency of lithology identification.

[0012] Further, the S4 comprises the following steps: S401, define a neighborhood set of lithology data nodes, the neighborhood set is used to represent the relationship between lithology data nodes, and is defined as follows: 1. For any two neighborhood subsets, if they are not the same subset, then their intersection is an empty set; 2. Lithology data nodes in the same neighborhood subset are considered to have the same lithology; 3. Lithology data nodes in the neighborhood set but not in the same neighborhood subset are considered to have different lithologies; before the first iteration starts, there is no neighborhood subset in the neighborhood set; S402: According to the query sequence Interact with experts and update the neighborhood set according to the lithology identification results given by experts, thereby optimizing the relationship structure of the neighbor diffusion network.

[0013] The beneficial effect of the above further solution is: by generating a query sequence , the more critical lithology data in the data set are handed over to experts for judgment, and the lithology of the key data is identified first; by generating and updating the neighborhood set of lithology data nodes, the expert's judgment is retained to guide the subsequent lithology identification process; compared with other lithology data, the key lithology data provides a more reliable reference for subsequent lithology identification. This step not only improves the accuracy of lithology identification, but also accelerates the convergence speed of the algorithm and improves efficiency.

[0014] Furthermore, the detailed steps of S6 are as follows: If the specified number of iterations has been reached, the lithology recognition result is directly output as the final recognition result; if the specified number of iterations has not been reached, the suspicious degree sequence is updated according to the lithology recognition result. , further generate key node sequences , and then further generate the query sequence Then, the neighborhood set of the lithology data nodes is updated based on the judgment given by the experts, and the annotation diffusion of the expert constraints is combined with the neighbor diffusion network until the identification results of all lithology data nodes are given; this process is iterated until the number of iterations reaches the specified number of iterations, and the final identification result is output.

[0015] The beneficial effect of the above further scheme is: by iteratively selecting the more critical lithology data nodes from the entire data node set, a query sequence is generated. Hand it over to experts for lithology identification, update the neighborhood set, and optimize the neighbor diffusion network structure; identify the lithology of other data nodes that are not directly identified by experts through the diffusion of expert constraints; at the same time, dynamically update the criticality of lithology data to generate high-quality query sequences , thereby ensuring the effectiveness of human-computer interaction and the accuracy of lithology identification; while reducing costs, it also improves the accuracy of lithology identification results.

[0016] Furthermore, the generation of the neighbor diffusion network includes the following steps: First, define the neighbor diffusion network ; Among them, the lithology data node set V Contains information corresponding to all logging data; edge set EIt includes the distance between lithology data nodes. There are edges between lithology data nodes, and the weight of the edge is calculated by the Euclidean distance between lithology data nodes. The neighbor diffusion network provides a medium for the subsequent annotation diffusion based on expert constraints. Then, the following operations are performed iteratively: a lithology data node is selected as a proxy node in each connected component, and each proxy node is connected to another proxy node that is closest to itself. When all proxy nodes are processed, proxy nodes are selected from each new connected component again until only one proxy node is left in the graph. After the iteration, only one connected component is left in the entire graph, and the neighbor diffusion network is obtained. It is worth noting that before the first iteration, there is no edge in the graph, and all lithology data nodes will be regarded as proxy nodes. The measurement indicator of whether a node is regarded as a proxy node is marked as , The value of is composed of the sum of two parts, one is the degree of the node; the other is the ratio of the sum of the weights of the adjacent edges of the point to the sum of the weights of the connected components in which it is located; The lithology data node with the largest value will be considered as the proxy node.

[0017] Furthermore, the representative sequence The generation of includes the following steps: On the basis of the neighbor diffusion network, the topological importance of the lithology data nodes is ranked in the following way: those lithology data nodes that have not been visited and are connected by the edge with the largest current weight are selected; first, the lithology data node with the largest degree is regarded as the lithology data node with the highest topological importance and is added to the representative sequence. Then, the nodes with the highest topological importance among the remaining lithology data nodes are added to the representative sequence in turn. Until the representative sequence All lithology data nodes are included.

[0018] Furthermore, the suspicious degree sequence The generation of includes the following steps: First, the suspiciousness of the lithology data node is related to the nearest The category to which the unit belongs; It is a parameter to be set, which represents the selection of the node closest to the lithology data node. nodes to measure their suspiciousness; secondly, Indicates this The lithology data node contains several classes in total ( ), neighboring nodes Belong to category The relationship is recorded as a node right The attributed value is expressed as , if the node Belongs to category ,So The value of is 1, otherwise The value of is 0; again, the node Nearest neighbor The lithology data nodes belong to the category The ratio is recorded as the neighbor attribution rate, expressed as , whose value is Node pairs The sum of the attribute values ​​is greater than the number of the nearest neighbor lithology data nodes , and its calculation formula is: In addition, the suspiciousness of the lithology data nodes is measured using Shannon entropy, expressed as , whose value is the product of the neighbor attribution rate and the logarithm of the neighbor attribution rate with base 2, summed for each class and then taking the inverse; the suspicious degree calculation formula is as follows: ; Finally, the suspiciousness of all lithology data nodes is sorted in descending order to obtain the suspiciousness sequence .

[0019] Furthermore, the key node sequence The generation of includes the following steps: First, the sequence length is specified as ; Then, select the key lithology data nodes. If the number of lithology data nodes with non-zero suspiciousness is greater than or equal to the sequence length, then select the suspiciousness sequence The front Nodes are used as key lithology data nodes to form a key node sequence If the number of lithology nodes with non-zero suspiciousness is less than the sequence length, the lithology data nodes with non-zero suspiciousness are added to the key node sequence in descending order of suspiciousness. Then, from the representative sequence Select the lithology data nodes with relatively high topological importance and 0 suspiciousness and add them to the key node sequence middle.

[0020] Further, the query sequence The generation of includes the following steps: First, introduce a variable , which records the key node sequence The currently visited node in ; in particular, during the first iteration, the key node sequence The first node in will be regarded as a new neighborhood subset and directly added to the neighborhood set; in other cases, the key node sequence is directly added to the The first lithology data node in is assigned to ; Secondly, the front of each neighborhood subset The nodes with the highest topological importance are selected when the length of the neighborhood subset is less than When , all nodes in the neighborhood subset are selected; among them, are artificially set parameters; finally, according to them and The distances of the corresponding lithology data nodes are sorted in ascending order to obtain the current The query sequence of the corresponding lithology data node .

[0021] Furthermore, the updating of the neighborhood set comprises the following steps: First, remember the key node sequence currently being processed The lithology data node in ,according to A query sequence is generated Subsequently, experts and query sequence Each node in the sequence is judged, if has the same lithology as the lithology data node in the neighborhood subset, then will be added to the neighborhood subset; if With query sequence If the lithology of any node in is different, then As a new neighborhood subset, add it to the neighborhood set; finally, end the Query.

[0022] Furthermore, the expert-constrained annotation diffusion on the neighbor diffusion network includes the following steps: First, for the neighborhood set formed after the query, a label is assigned to each neighborhood subset to distinguish the lithology data nodes in different neighborhood subsets; then, for each lithology data node in the first neighborhood subset, the nodes directly connected to it but not in the neighborhood set are assigned the label of the neighborhood subset, until all the lithology data nodes in the first neighborhood subset and the nodes directly connected to them are assigned labels, and then the second neighborhood subset is diffused until the diffusion of the last neighborhood subset is completed; the diffusion will stop after the lithology of all lithology data nodes has been identified, and the end of the diffusion marks the generation of the lithology identification results of this round of iteration. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present invention are described in detail below, but it should be clear that the embodiments described here are only a part, and the present invention is not limited to the scope of the specific implementation methods. For ordinary technicians in this field, changes to the embodiments herein that are limited within the claims belong to the protection scope of the present invention.

[0025] like Figure 1 As shown, the present invention provides a method for identifying underground reservoir lithology based on active learning technology, and its implementation method is as follows: S1. Collection and collation of original logging data and conventional logging curve data, and preprocessing of existing lithology data, the implementation method is as follows: S101. Obtain lithology data, including well diameter, gamma ray, natural potential, resistivity, acoustic wave, density, and photoelectric absorption index; S102, preprocessing the lithology data, filling in missing values, and extracting features; S2. Generate a neighbor diffusion network based on the preprocessed lithology data. The implementation method is as follows: S201, taking all the lithology data after preprocessing as nodes in a neighbor diffusion network; S202, selecting a proxy node from the connected components formed by each lithology data node, and merging the connected components; S3. Dynamically update the criticality of lithology data nodes based on the neighbor diffusion network, and then obtain the critical node sequence And the query sequence ; The implementation method is as follows: S301. Calculate the topological importance of all nodes and generate representative sequences ; S302: Calculate the suspiciousness of all nodes and generate a suspiciousness sequence ; S303, based on representative sequences and suspicious degree sequence Generate key node sequence ; S304, according to the key node sequence Generate query sequence ; S4. Based on the expert's judgment, the neighborhood set of lithology data nodes is updated, and a small number of lithology data nodes are identified and the relationship structure of the neighbor diffusion network is optimized; S401, define a neighborhood set of lithology data nodes, the neighborhood set is used to represent the relationship between lithology data nodes, and is defined as follows: 1. For any two neighborhood subsets, if they are not the same subset, then their intersection is an empty set; 2. Lithology data nodes in the same neighborhood subset are considered to have the same lithology; 3. Lithology data nodes in the neighborhood set but not in the same neighborhood subset are considered to have different lithologies; before the first iteration starts, there is no neighborhood subset in the neighborhood set; S402: According to the query sequence Interact with experts and update the neighborhood set according to the lithology identification results given by experts, thereby optimizing the relationship structure of the neighbor diffusion network; S5. Perform annotation diffusion based on expert constraints on the nearest neighbor diffusion network to obtain the lithology identification result of the current round; S6, iteratively optimize the neighbor diffusion network relationship structure, stop iteration after reaching the specified number of iterations, and output the recognition result. The specific steps are as follows: If the specified number of iterations has been reached, the lithology recognition result is directly output as the final recognition result; if the specified number of iterations has not been reached, the suspicious degree sequence is updated according to the lithology recognition result. , further generate key node sequences , and then further generate the query sequence Then, the neighborhood set of the lithology data nodes is updated based on the judgment given by the experts, and the annotation diffusion of the expert constraints is combined with the neighbor diffusion network until the identification results of all lithology data nodes are given; this process is iterated until the number of iterations reaches the specified number of iterations, and the final identification result is output.

[0026] In this embodiment, the present invention combines active learning technology to identify the lithology of underground reservoirs; specifically, firstly, the original logging data is obtained and preprocessed, wherein the original logging data includes well diameter, gamma ray, natural potential, resistivity, acoustic wave, density, and photoelectric absorption index; then, by generating a neighbor diffusion network, the association structure between the lithology data is established, which provides a medium for subsequent diffusion based on expert annotation, and the criticality of the lithology data is measured by measuring the importance and suspicion of the topological structure, thereby generating a query sequence and interact with experts; it is worth mentioning that this interaction is not a one-time thing. During the entire implementation process, the model will continuously select the most critical lithology data nodes for experts to judge and perform self-optimization, thereby maximizing the effectiveness of the query, and improving the accuracy of recognition by continuously updating the criticality of the data and the neighborhood set; and giving a current recognition result in each round; until the number of iterations reaches the specified number of iterations, a final recognition result of a lithology data node is given.

[0027] In this embodiment, the logging data includes wellbore diameter, gamma ray, natural potential, resistivity, acoustic wave, density, and photoelectric absorption index.

[0028] In this embodiment, logging data is obtained from each wellhead, missing values ​​are filled in the original logging data, and each feature is regarded as a dimension and each lithology data is regarded as a node as the initial node set of the neighbor diffusion network.

[0029] In this embodiment, the logging data matrix is ​​generated : ; in, represents the well diameter measured in the nth depth interval, represents the intensity of gamma rays measured in the nth depth interval, represents the natural potential difference measured in the nth depth interval, represents the deep lateral resistivity measured in the nth depth interval, represents the shallow lateral resistivity measured in the nth depth interval, It represents the acoustic time difference measured in the nth depth interval, represents the density measured in the nth depth interval, It represents the photoelectric absorption index measured in the nth depth interval. Each row represents a depth, and the data measured in the nth depth interval are located in the nth row. The elements in the matrix are specific values, and n represents the nth depth interval.

[0030] In this embodiment, the generation of the neighbor diffusion network includes the following steps: First, define the neighbor diffusion network ; Among them, the lithology data node set V Contains information corresponding to all logging data; edge set E It includes the distance between lithology data nodes. There are edges between lithology data nodes, and the weight of the edge is calculated by the Euclidean distance between lithology data nodes. The neighbor diffusion network provides a medium for the subsequent annotation diffusion based on expert constraints. Then, the following operations are performed iteratively: a lithology data node is selected as a proxy node in each connected component, and each proxy node is connected to another proxy node that is closest to itself. When all proxy nodes are processed, proxy nodes are selected from each new connected component again until only one proxy node is left in the graph. After the iteration, only one connected component is left in the entire graph, and the neighbor diffusion network is obtained. It is worth noting that before the first iteration, there is no edge in the graph, and all lithology data nodes will be regarded as proxy nodes. The measurement index of whether a node is regarded as a proxy node is marked as ; The value of is composed of the sum of two parts, one is the degree of the node; the other is the ratio of the sum of the weights of the adjacent edges of the point to the sum of the weights of the connected components in which it is located; The lithology data node with the largest value will be considered as the proxy node.

[0031] In this example, the representative sequence The generation of the representative sequence It is used to measure the topological importance of nodes in the graph; based on the neighbor diffusion network, the topological importance of lithology data nodes is sorted in the following way: those lithology data nodes that have not been visited and are connected by the edge with the largest current weight are selected; first, the lithology data node with the largest degree is regarded as the lithology data node with the highest topological importance and is added to the representative sequence Then, the nodes with the highest topological importance among the remaining lithology data nodes are added to the representative sequence in turn. Until the representative sequence All lithology data nodes are included.

[0032] In this embodiment, the suspicious degree sequence The generation of includes the following steps: First, the suspiciousness of the lithology data node is compared with the nearest The category to which the unit belongs; It is a parameter to be set, which represents the selection of the node closest to the lithology data node. nodes to measure their suspiciousness; secondly, Indicates this The lithology data node contains several classes in total ( ), neighboring nodes Belong to category The relationship is recorded as a node right The attributed value is expressed as , if the node Belongs to category ,So The value of is 1, otherwise The value of is 0; again, the node Nearest neighbor The lithology data nodes belong to the category The ratio is recorded as the neighbor attribution rate, expressed as , whose value is Node pairs The sum of the attribute values ​​is greater than the number of the nearest neighbor lithology data nodes , and its calculation formula is: In addition, the suspiciousness of the lithology data nodes is measured using Shannon entropy, expressed as , whose value is the product of the neighbor attribution rate and the logarithm of the neighbor attribution rate with base 2, summed for each class and then taking the inverse; the suspicious degree calculation formula is as follows: ; Finally, the suspiciousness of all lithology data nodes is sorted in descending order to obtain the suspiciousness sequence .

[0033] In this embodiment, the key node sequence The generation of includes the following steps: First, the sequence length is specified as ,Then, select the key lithology data nodes. If the number of lithology data nodes with non-zero suspiciousness is greater than or equal to the sequence length, then select the suspiciousness sequence The front Nodes are used as key lithology data nodes to form a key node sequence If the number of lithology nodes with non-zero suspiciousness is less than the sequence length, the lithology data nodes with non-zero suspiciousness are added to the key node sequence in descending order of suspiciousness. Then, from the representative sequence Select the lithology data nodes with relatively high topological importance and 0 suspiciousness and add them to the key node sequence middle.

[0034] In this embodiment, according to the key node sequence For query sequence The generation of includes the following steps: First, introduce a variable , which records the key node sequence The currently visited node in ; in particular, during the first iteration, the key node sequence The first node in will be regarded as a new neighborhood subset and directly added to the neighborhood set; in other cases, the key node sequence is directly added to the The first lithology data node in is assigned to ; Secondly, the front of each neighborhood subset The nodes with the highest topological importance are selected; when the length of the neighborhood subset is less than When , all nodes in the neighborhood subset are selected; among them, are artificially set parameters; finally, according to them and The distances of the corresponding lithology data nodes are sorted in ascending order to obtain the current The query sequence of the corresponding lithology data node .

[0035] In this embodiment, the updating of the neighborhood set includes the following steps: first, in the first iteration, for the first node, since there is no neighborhood subset in the neighborhood set, it is regarded as a neighborhood subset and added to the neighborhood set; then, for the second node, it is asked whether the lithology is the same as that of the first lithology data node. If the expert gives an affirmative result, the second node is added to the neighborhood subset formed by the first lithology data node. If the expert gives a negative result, the second lithology data node will also form a new neighborhood subset by itself and be added to the neighborhood set; finally, the query is terminated until the expert gives a result that the queried node and the query node belong to the same category or all query nodes have been visited. If the query sequence Before the query is exhausted, the expert has given that the queried node belongs to the same category as a query node, so the query about the queried node ends, and the query sequence is regenerated for the next queried node. ; If the expert believes that no query node belongs to the same category as the current queried node, the queried node will generate a new neighborhood subset containing only itself and be added to the neighborhood set.

[0036] In this embodiment, the method of performing expert-constrained label diffusion on the neighbor diffusion network is as follows: first, for the neighborhood set formed after the query, a label is assigned to each neighborhood subset therein to distinguish the lithology data nodes in different neighborhood subsets; then, for each lithology data node in the first neighborhood subset, the nodes directly connected to it but not in the neighborhood set are assigned the label of the neighborhood subset, until all the lithology data nodes in the first neighborhood subset and the nodes directly connected to them are assigned labels, and then the second neighborhood subset is diffused until the diffusion of the last neighborhood subset is completed; the diffusion will stop after the lithology of all lithology data nodes is identified, and the end of the diffusion marks the generation of the lithology identification results of this round of iteration.

[0037] In summary, compared with the prior art, the underground reservoir lithology identification method based on active learning technology disclosed in the present invention has the following beneficial effects: (1) Comprehensively measure data characteristics: key node sequence The query sequence is further generated by comprehensively evaluating the topological importance and suspiciousness of the lithology nodes. , while ensuring the maximum effect of human-computer interaction, it also ensures a comprehensive measurement of data features; (2) Reduce costs: Compared with existing methods, experts need to perform fewer annotations, which effectively reduces costs; (3) Improve recognition accuracy and achieve stable results: query sequence is through the key node sequence It is generated by further screening, effectively avoiding the processing of noise data; it iteratively re-measures various indicators of the data to ensure its stability; and it improves the recognition accuracy by continuously optimizing the neighbor diffusion network structure.

Claims

1. A method for identifying underground reservoir lithology based on active learning technology, characterized in that: The following steps are involved: S1. Collection and collation of original logging data and conventional logging curve data, and preprocessing of existing lithology data; S2, generating a neighbor diffusion network based on the preprocessed lithology data; S3. Dynamically update the criticality of lithology data nodes based on the neighbor diffusion network, and then obtain the critical node sequence And the query sequence ; S4. Based on the expert's judgment, the neighborhood set of lithology data nodes is updated, and a small number of lithology data nodes are identified and the relationship structure of the neighbor diffusion network is optimized; S5. Perform annotation diffusion based on expert constraints on the nearest neighbor diffusion network to obtain the lithology identification result of the current round; S6. Iteratively optimize the neighbor diffusion network relationship structure, stop iterating after reaching the specified number of iterations, and output the recognition result.

2. The underground reservoir lithology identification method based on active learning technology according to claim 1 is characterized in that: The S1 comprises the following steps: S101. Obtain lithology data, including well diameter, gamma ray, natural potential, resistivity, acoustic wave, density, and photoelectric absorption index; S102, preprocessing the lithology data; filling in missing values ​​and extracting features.

3. The underground reservoir lithology identification method based on active learning technology according to claim 1 is characterized in that: The S2 comprises the following steps: S201, taking all the lithology data after preprocessing as nodes in a neighbor diffusion network; S202: Select a proxy node from the connected components formed by each lithology data node, and merge the connected components.

4. The underground reservoir lithology identification method based on active learning technology according to claim 1 is characterized in that: The S3 comprises the following steps: S301. Calculate the topological importance of all nodes and generate representative sequences ; S302: Calculate the suspiciousness of all nodes and generate a suspiciousness sequence ; S303, based on representative sequences and suspicious degree sequence Generate key node sequence ; S304, according to the key node sequence Generate query sequence .

5. The underground reservoir lithology identification method based on active learning technology according to claim 1 is characterized in that: The S4 comprises the following steps: S401, define a neighborhood set of lithology data nodes, the neighborhood set is used to represent the relationship between lithology data nodes, and is defined as follows:

1. For any two neighborhood subsets, if they are not the same subset, then their intersection is an empty set; 2. Lithology data nodes in the same neighborhood subset are considered to have the same lithology; 3. Lithology data nodes in the neighborhood set but not in the same neighborhood subset are considered to have different lithologies; before the first iteration starts, there is no neighborhood subset in the neighborhood set; S402: According to the query sequence Interact with experts and update the neighborhood set according to the lithology identification results given by experts, thereby optimizing the relationship structure of the neighbor diffusion network.

6. The underground reservoir lithology identification method based on active learning technology according to claim 1 is characterized in that: The detailed steps of S6 are as follows: If the specified number of iterations has been reached, the lithology recognition result is directly output as the final recognition result; if the specified number of iterations has not been reached, the suspicious degree sequence is updated according to the lithology recognition result. , further generate key node sequences , and then further generate the query sequence Then, the neighborhood set of the lithology data nodes is updated based on the judgment given by the experts, and the annotation diffusion of the expert constraints is combined with the neighbor diffusion network until the identification results of all lithology data nodes are given; this process is iterated until the number of iterations reaches the specified number of iterations, and the final identification result is output.

7. The underground reservoir lithology identification method based on active learning technology according to claim 3 is characterized in that: The generation of the neighbor diffusion network includes the following steps: First, define the neighbor diffusion network ; Among them, the lithology data node set V Contains information corresponding to all logging data; edge set E It includes the distance between lithology data nodes. There are edges between lithology data nodes, and the weight of the edge is calculated by the Euclidean distance between lithology data nodes. The neighbor diffusion network provides a medium for the subsequent annotation diffusion based on expert constraints. Then, the following operations are performed iteratively: a lithology data node is selected as a proxy node in each connected component, and each proxy node is connected to another proxy node that is closest to itself. When all proxy nodes are processed, proxy nodes are selected from each new connected component again until only one proxy node is left in the graph. After the iteration, only one connected component is left in the entire graph, and the neighbor diffusion network is obtained. It is worth noting that before the first iteration, there is no edge in the graph, and all lithology data nodes will be regarded as proxy nodes. The measurement index of whether a node is regarded as a proxy node is marked as ; The value of is composed of the sum of two parts, one is the degree of the node; the other is the ratio of the sum of the weights of the adjacent edges of the point to the sum of the weights of the connected components in which it is located; The lithology data node with the largest value will be considered as the proxy node.

8. The underground reservoir lithology identification method based on active learning technology according to claim 4 is characterized in that: The representative sequence The generation of includes the following steps: On the basis of the neighbor diffusion network, the topological importance of the lithology data nodes is ranked in the following way: those lithology data nodes that have not been visited and are connected by the edge with the largest current weight are selected; first, the lithology data node with the largest degree is regarded as the lithology data node with the highest topological importance and is added to the representative sequence. Then, the nodes with the highest topological importance among the remaining lithology data nodes are added to the representative sequence in turn. Until the representative sequence All lithology data nodes are included.

9. The underground reservoir lithology identification method based on active learning technology according to claim 4, characterized in that: The suspicious degree sequence The generation of includes the following steps: First, the suspiciousness of the lithology data node is related to the nearest The category to which the unit belongs; It is a parameter to be set, which represents the selection of the node closest to the lithology data node. nodes to measure their suspiciousness; secondly, Indicates this The lithology data node contains several classes in total ( ), neighboring nodes Belong to category The relationship is recorded as a node right The attributed value is expressed as , if the node Belongs to category ,So The value of is 1, otherwise The value of is 0; again, the node Nearest neighbor The lithology data nodes belong to the category The ratio is recorded as the neighbor attribution rate, expressed as , whose value is Node pairs The sum of the attribute values ​​is greater than the number of the nearest neighbor lithology data nodes , and its calculation formula is: In addition, the suspiciousness of the lithology data nodes is measured using Shannon entropy, expressed as , whose value is the product of the neighbor attribution rate and the logarithm of the neighbor attribution rate with base 2, summed for each class and then taking the inverse; the suspicious degree calculation formula is as follows: ； Finally, the suspiciousness of all lithology data nodes is sorted in descending order to obtain the suspiciousness sequence .

10. The underground reservoir lithology identification method based on active learning technology according to claim 4, characterized in that: The key node sequence Generation The following steps are involved: First, the sequence length is specified as ; Then, select the key lithology data nodes. If the number of lithology data nodes with non-zero suspiciousness is greater than or equal to the sequence length, then select the suspiciousness sequence The front Nodes are used as key lithology data nodes to form a key node sequence ; If the number of lithology nodes with non-zero suspiciousness is less than the sequence length, the lithology data nodes with non-zero suspiciousness are added to the key node sequence in descending order of suspiciousness. Then, from the representative sequence Select the lithology data nodes with relatively high topological importance and 0 suspiciousness and add them to the key node sequence middle.

11. The underground reservoir lithology identification method based on active learning technology according to claim 4, characterized in that: The query sequence The generation of includes the following steps: First, introduce a variable , used to record the key node sequence The currently visited node in ; in particular, during the first iteration, the key node sequence The first node in will be regarded as a new neighborhood subset and directly added to the neighborhood set; in other cases, the key node sequence is directly added to the The first lithology data node in is assigned to ; Secondly, the front of each neighborhood subset The nodes with the highest topological importance are selected; when the length of the neighborhood subset is less than When , all nodes in the neighborhood subset are selected; among them, are artificially set parameters; finally, according to them and The distances of the corresponding lithology data nodes are sorted in ascending order to obtain the current The query sequence of the corresponding lithology data node .

12. The underground reservoir lithology identification method based on active learning technology according to claim 5, characterized in that: The updating of the neighborhood set comprises the following steps: First, remember the key node sequence currently being processed The lithology data node in ,according to A query sequence is generated Subsequently, experts and query sequence Each node in the sequence is judged, if has the same lithology as the lithology data node in the neighborhood subset, then will be added to the neighborhood subset; if With query sequence If the lithology of any node in is different, then As a new neighborhood subset, add it to the neighborhood set; finally, end the Query.

13. The underground reservoir lithology identification method based on active learning technology according to claim 1, characterized in that: The method of performing expert-constrained annotation diffusion on a neighbor diffusion network includes the following steps: First, for the neighborhood set formed after the query, a label is assigned to each neighborhood subset to distinguish the lithology data nodes in different neighborhood subsets; then, for each lithology data node in the first neighborhood subset, the nodes directly connected to it but not in the neighborhood set are assigned the label of the neighborhood subset, until all the lithology data nodes in the first neighborhood subset and the nodes directly connected to them are assigned labels, and then the second neighborhood subset is diffused until the diffusion of the last neighborhood subset is completed; the diffusion will stop after the lithology of all lithology data nodes has been identified, and the end of the diffusion marks the generation of the lithology identification results of this round of iteration.

Images