Multi-mode intelligent while-drilling lithology identification method based on prior constraint and expert interaction

Abstract

The invention discloses a multi-modal intelligent while-drilling lithology identification method based on prior constraint and expert interaction. The method comprises the following steps: constructing a multi-modal data set; constructing and training a multi-modal rock debris classification model; establishing a constraint list; and S3, outputting an original probability distribution vector P by using a multi-modal rock debris classification model, generating a mask vector M through the constraint list established in the step S3, and calculating a final output probability Pfinal. According to the method, by fusing multi-modal data such as images, logging and elements, cross verification and supplement of information are achieved, and the accuracy and reliability of complex lithology recognition are improved; priori knowledge such as geological background, expert experience and the like is used as hard constraint conditions of the model, so that an anti-geological prediction result is fundamentally eradicated, and a decision is enabled to conform to a geological law; according to the method and the system, implicit knowledge of experts is dominated and modeled, organic combination of artificial intelligence and expert intelligence is realized, and on the premise of ensuring accuracy, manpower is greatly liberated, and timeliness of well logging is improved.

Description

Technical Field

[0001] This invention belongs to the field of geological logging technology in oil and gas exploration, specifically involving a multimodal intelligent drilling lithology identification method based on prior constraints and expert interaction. Background Technology

[0002] Cuttings logging while drilling is the most direct and economical technique for understanding the lithology of downhole formations during oil and gas exploration and development. By analyzing the cuttings returned to the surface during drilling, key geological information such as the lithology and hydrocarbon potential of the encountered formations can be determined in real time, providing a basis for decision-making regarding geological steering, reservoir evaluation, and drilling safety. Its application spans the entire drilling process, from exploration wells to appraisal wells and development wells, and it plays an irreplaceable role in the efficient and low-cost discovery and development of oil and gas resources.

[0003] Traditional drilling lithology identification relies primarily on logging engineers' personal experience to observe the color, structure, mineral composition, luster, and hardness of cuttings, and then name the lithology by referring to standard charts. The accuracy of lithology identification is directly related to the logging engineer's experience level and condition (such as fatigue level). Different personnel may name the same cuttings sample differently, lacking objectivity and consistency.

[0004] Qualified cuttings logging engineers require extensive professional training and practical experience. Furthermore, the manual description and naming process is time-consuming and cannot meet the demands of high-speed, high-efficiency modern drilling.

[0005] For strata with complex lithological assemblages and diverse mineral compositions, such as igneous rocks, metamorphic rocks, or fine-grained sedimentary rocks, even experienced experts cannot guarantee the accuracy of drilling cuttings identification. For example, in volcanic strata, andesite, basalt, and tuff have similar characteristics in terms of cuttings morphology, making them easily confused and leading to errors in stratigraphic correlation and distorted reservoir evaluation.

[0006] With the development of artificial intelligence technology, especially the successful application of deep learning in image recognition, some intelligent lithology identification methods based on rock cuttings images have emerged. These methods typically employ models such as convolutional neural networks (CNNs) to automatically classify rock cuttings photos, replacing manual identification. However, current AI-powered logging technologies generally suffer from the following bottlenecks: 1) Limited Data Modality and Inadequate Information Utilization: Most methods rely solely on two-dimensional image information of rock cuttings, neglecting other valuable data reflecting the physical and chemical properties of rocks, such as logging-while-drilling data (natural gamma ray GR, resistivity, neutron density, etc.), well-side seismic traces, X-ray fluorescence (XRF) elemental analysis data, and X-ray diffraction (XRD) mineral content data. These multimodal data contain characteristic information of rocks in different dimensions. The application of a single modality (image) results in the loss of a large amount of crucial information, limiting the accuracy and robustness of identification.

[0007] 2) Lack of geological interpretability, becoming a "black box" model: Traditional deep learning models are typically end-to-end "input-output" systems, and their decision-making processes are difficult to interpret. They cannot tell geological experts which features (such as specific mineral grains or structures) the model bases its judgments on. This "black box" characteristic does not meet the geological research requirements for exploring causes and processes, making it difficult for geological experts to trust, verify, and correct the model's identification results.

[0008] 3) Ignoring geological context constraints can easily lead to "counter-geological" errors: Rock formation is not random but follows strict geological laws. For example, igneous rocks cannot occur in a known marine sedimentary environment. Existing AI models often lack mechanisms to incorporate this geological context knowledge, and their predictions may contradict regional geological patterns, resulting in basic "counterintuitive" errors.

[0009] 4) Lack of effective expert interaction mechanisms: Existing models are mostly trained offline and predicted online, making it difficult to conveniently incorporate expert prior knowledge for real-time intervention and correction during application. When applied to new exploration blocks, the inability to quickly adjust and adapt using expert experience results in poor model generalization ability. Summary of the Invention

[0010] This invention is proposed to solve the problems existing in the prior art, and its purpose is to provide a multimodal intelligent drilling lithology identification method based on prior constraints and expert interaction.

[0011] This invention is achieved through the following technical solution: A multimodal intelligent drilling-while-drilling lithology identification method based on prior constraints and expert interaction includes the following steps: S1. Construct a multimodal dataset; S2. Construction and training of a multimodal rock cuttings classification model; S3. Create a constraint list; S4. Using the multimodal rock cuttings classification model, output the original probability distribution vector P. Generate a mask vector M using the constraint list established in step S3, and calculate the final output probability P. final .

[0012] In the above technical solution, step S1 specifically includes: S11. Synchronously collect modal data for each rock cutting sample; S12. Take core samples from the well section corresponding to the cuttings sample, prepare polarized thin sections, determine the final lithological name through cross-identification, and set the lithological name determined by cross-identification as the gold standard label dataset. S13. A training set is constructed using a random masking strategy. For each batch of training data, one or more modal data are randomly set to zero or replaced with invalid values ​​with a set probability.

[0013] In the above technical solution, the modal data includes unstructured modal data and structured modal data; The unstructured modal data is rock cutting photographs; the rock cutting photographs are high-resolution digital images of cleaned rock cutting samples taken with an industrial camera under standard light sources. The structured modal data includes logging-while-drilling data, well-side seismic trace data, and element / mineral content data. The logging-while-drilling data includes GR, SP, RT, AC, DEN, and CNL corresponding to the depth of the cuttings sample. The well-side seismic trace data is three-dimensional seismic trace data near the depth point of the cuttings sample. The element / mineral content data is obtained by rapid analysis of the relative contents of major elements and minerals in the cuttings sample using XRF and XRD.

[0014] In the above technical solution, step S1 further includes: S14. Construct a dataset of random lithological combinations that supports prior constraints.

[0015] In the above technical solution, step S2 specifically includes: S21. Using a general convolutional neural network as the backbone network, process the input rock debris photos to extract high-dimensional feature vectors V from the unstructured data. img All structured data are concatenated into a one-dimensional vector, and then a high-dimensional feature vector V is extracted from the structured data using a multilayer perceptron network. struct ; S22, using high-dimensional feature vectors V from unstructured data img and structured data high-dimensional feature vector V struct Calculate the correlation weights between the features of unstructured and structured data modalities, and generate a fused feature vector V that integrates information from both modalities. fused ; S23, Utilizing the fused feature vector V fused The probability distribution vector is output through the Softmax activation function; The expression for the probability distribution vector is: P=[p1,p2,...,p n ] In the formula: n is the total number of lithological categories, p i The representative model predicts the probability that the rock fragment sample belongs to the i-th lithology, i=1,...,n; S24. Using the multimodal dataset constructed in step S1, and with the cross-entropy loss function as the objective function, optimize and train the parameters of the entire model through the backpropagation algorithm.

[0016] In the above technical solution, the unstructured high-dimensional feature vector V img This includes information on the color, texture, and grain morphology of rock debris photographs.

[0017] In the above technical solution, the step of using high-dimensional feature vector V of unstructured data... img and structured data high-dimensional feature vector V struct Methods for achieving multimodal data fusion include: using cross-modal attention mechanisms, directly feeding feature concatenation into an MLP, or using multi-head attention mechanisms for finer-grained feature interaction.

[0018] In the above technical solution, the constraint list includes prior knowledge constraints and expert knowledge constraints; the prior knowledge constraints are geological laws.

[0019] In the above technical solution, the method for establishing the constraint list is as follows: using a large language model to extract the natural language of the input text description, generating a list of possible lithological combinations, and thus completing the establishment of the constraint list.

[0020] In the above technical solution, the final output probability P final The calculation formula is: P final =Softmax(P⊙M).

[0021] The beneficial effects of this invention are: This invention provides a multimodal intelligent drilling lithology identification method based on prior constraints and expert interaction. By fusing multimodal data such as images, well logging, and elements, it achieves cross-validation and supplementation of information, significantly improving the accuracy and reliability of complex lithology identification. It uses prior knowledge such as geological background and expert experience as hard constraints on the model, fundamentally eliminating "anti-geological" predictions and ensuring that the AI ​​model's decisions conform to geological laws. It provides a convenient expert interaction interface, making the implicit knowledge of experts explicit and modeled, achieving an organic combination of artificial intelligence and expert wisdom. While ensuring accuracy, it greatly liberates manpower and improves the timeliness of logging. This invention has significant application and promotion value for promoting the intelligent, standardized, and efficient development of logging technology, and can provide more reliable technical support for the accurate identification and efficient development of oil and gas reservoirs. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] For those skilled in the art, other related figures can be obtained from the above figures without any creative effort. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0025] Example 1 like Figure 1 As shown, a multimodal intelligent drilling-while-drilling lithology identification method based on prior constraints and expert interaction includes the following steps: S1. Construct a multimodal dataset: S11. Synchronously collect modal data for each rock cutting sample; The modal data includes unstructured modal data and structured modal data; The unstructured modal data is rock cutting photographs; the rock cutting photographs are high-resolution digital images of cleaned rock cutting samples taken with an industrial camera under standard light sources. The structured modal data includes logging-while-drilling data, well-side seismic trace data, and element / mineral content data; The logging-while-drilling data includes logging curve values ​​such as GR (natural gamma), SP (spontaneous potential), RT (deep resistivity), AC (acoustic transit time), DEN (density), and CNL (neutrons) corresponding to the depth of the cuttings sample. The well-side seismic trace data are three-dimensional seismic trace data near the depth point of the rock cuttings sample; The elemental / mineral content data were obtained by rapid analysis of the relative contents of major elements and main minerals in the rock fragment samples using XRF (X-ray fluorescence spectrometry) and XRD (X-ray diffraction).

[0026] Unstructured modal data and structured modal data together constitute a multimodal dataset; S12. Take core samples from the well section corresponding to the cuttings sample, prepare polarized thin sections, determine the final lithological name through cross-identification, and set the lithological name determined by cross-identification as the gold standard label of the multimodal dataset. The method for preparing the polarizing thin film is in accordance with SY / T 5913-2021 Rock Slicing Method; The specific method for determining the final lithological name through cross-analysis is as follows: the lithological name is determined by combining thin section, XRF, and XRD. S13. A training set is constructed using a random masking strategy. For each batch of training data, one or more modal data are randomly set to zero or replaced with invalid values ​​with a set probability. The training data originates from a multimodal dataset composed of both unstructured modal data and structured modal data. For example, for a particular sample, during training, only its "cuttings photos" and "GR values" might be used, while the "resistivity" and "XRF data" are masked. The direct technical effect of this feature is that it allows the model to avoid over-reliance on any single data source during training, enabling it to learn to infer even with incomplete data. Thus, in practical applications, even if the user can only provide data for some modalities (e.g., a well section lacks XRF data), the model can still work robustly. S14. Construct a dataset of random lithological combinations that supports prior constraints; For example, if a stratum is known to be a set of marine-continental transitional sedimentary facies, and its possible lithological combinations are sandstone, siltstone, mudstone, and coal seams, then a subset of the dataset containing only these four lithologies can be constructed and an "constraint label" can be attached. This allows the model to learn this lithological combination relationship during training, and to match the "constraint label" with prior knowledge constraints and expert interaction information to achieve "hard constraints".

[0027] S2. Construction and training of a multimodal rock cuttings classification model: S21. Using a general convolutional neural network as the backbone network, process the input rock debris photos to extract high-dimensional feature vectors V from the unstructured data. img ; The unstructured high-dimensional feature vector V img This includes information on the color, texture, and grain morphology of the rock debris photographs; All structured modal data are concatenated into a one-dimensional vector, and then a high-dimensional feature vector V is extracted from the structured data using a multilayer perceptron network. struct ; Through the high-dimensional feature vector V of unstructured data img and structured data high-dimensional feature vector V struct The correlation weights between unstructured and structured modal data are dynamically calculated, enabling the model to autonomously learn how to focus on information from one modal data based on another, generating a fused feature vector V that integrates both unstructured and structured modal data. fused To achieve multimodal data fusion; The high-dimensional feature vector V of unstructured data img and structured data high-dimensional feature vector V struct Methods for achieving multimodal data fusion include: using cross-modal attention mechanisms, using feature concatenation and then directly feeding it into an MLP, or using multi-head attention mechanisms for finer-grained feature interaction. All three methods are conventional techniques in this field.

[0028] S22, Fuse the feature vector V fused The input is fed into a fully connected layer and activated by the Softmax function, ultimately outputting a probability distribution vector (using the fused feature vector V). fused (The probability distribution vector is output through the Softmax activation function). The expression for the probability distribution vector is: P=[p1,p2,...,p n ] In the formula: n is the total number of lithological categories, p i The representative model predicts the probability that the rock fragment sample belongs to the i-th lithology, i=1,...,n; S23. Using the multimodal dataset constructed in step S1, with the cross-entropy loss function as the objective function, optimize and train the parameters of the probability distribution vector through the backpropagation algorithm; In addition to the "hard constraint" of the component in step S14 (setting the probability of impossible lithology to zero), prior knowledge can also be added as a regularization term to the cross-entropy loss function to guide the model to learn in a direction that conforms to geological laws during the training phase, thus achieving "soft constraint"; that is, adding "prior knowledge" to the cross-entropy loss function. S3. Create a constraint list; The constraint list includes prior knowledge constraints and expert knowledge constraints; the prior knowledge constraints are geological laws. The constraint list is established by parsing the natural language textual description of the geological background of the work area input by experts using a large language model (LLM), extracting the lithological information, and generating a list of possible lithological combinations to complete the constraint list establishment.

[0029] For example: "This well section is located in the delta front subfacies of the XX depression, mainly developing underwater distributary channel sand bodies and sheet sands, with a small amount of interdistributary bay mudstone." The LLM can parse this natural language, automatically extract the lithological information contained therein, and generate a list of possible lithological combinations, such as ["sandstone", "siltstone", "mudstone"].

[0030] S4. Using the multimodal rock cuttings classification model, output the original probability distribution vector P. Generate a mask vector M using the constraint list established in step S3. For lithologies present in the constraint list, the corresponding position in the mask vector is 1; otherwise, it is 0. Calculate the final output probability P. final The final output probability P final It represents the probability of different lithologies, and the lithology with the highest probability is usually taken as the identification result.

[0031] The final output probability P final The calculation formula is: P final =Softmax(P⊙M).

[0032] This invention constructs a flexible lithology identification framework based on multimodal data input. Existing technologies are mostly based on single cuttings images or fixed combinations of several well logging data. This invention proposes a flexible framework that supports input combinations of any one or more types of data, including cuttings photographs, various well logging data, seismic data, and elemental / mineral content data. This greatly enhances the applicability and robustness of the method. In actual operations, data is often incomplete; this invention can maximize the use of all available data in the field to make the most reliable judgments.

[0033] This invention employs a human-machine collaborative mechanism with hard constraints based on prior knowledge. Existing AI logging technologies are purely data-driven and lack effective ways to incorporate geological principles and expert prior knowledge. This invention innovatively proposes a "hard constraint" mechanism, whereby the model's final output must strictly conform to the lithological range set by experts. This invention fundamentally solves the problem of potential "anti-geological" errors in AI models, ensuring that the AI's output aligns with the cognitive logic of geologists, significantly improving the scientific rigor and credibility of the results. This is a crucial step from "usable" to "reliable and easy to use."

[0034] This invention utilizes Large Language Models (LLMs) to automatically extract lithological constraints from geological natural language descriptions, providing a novel and efficient way to facilitate the input of expert knowledge. Traditional human-computer interaction requires experts to manually select or input parameters, while this invention allows experts to describe using their most familiar natural language. This invention significantly lowers the barrier to entry, enabling expert knowledge to be more easily and quickly integrated into AI decision-making processes, thus improving the intelligence and efficiency of the interaction.

[0035] The applicant declares that the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention fall within the protection and disclosure scope of the present invention.

Claims

1. A method for priori constraint and expert interaction of multi-modal intelligent LWD lithology identification, characterized in that: The method comprises the following steps: S1, constructing a multi-modal data set; S2, constructing and training a multi-modal cutting classification model; S3, establishing a constraint list; S4, using the multi-modal cutting classification model to output the original probability distribution vector P, and using the constraint list established in step S3 to generate a mask vector M, and calculating the final output probability P final .

2. The method of claim 1, wherein the method further comprises: The step S1 is specifically: S11, for each cutting sample, synchronously collecting modal data; S12, sampling the core of the well section corresponding to the cutting sample, making a polarized thin section, determining the final lithology name through cross identification, and setting the lithology name determined by cross identification as the gold standard label data set; S13, adopting a random mask strategy to construct a training set, and for each batch of training data, randomly setting one or more modal data to zero or replacing it with an invalid value with a set probability.

3. The method of claim 2, wherein the prior constraints and expert interaction multi-modal intelligent LWD lithology identification method is characterized by: The modal data includes unstructured modal data and structured modal data; The unstructured modal data is a cutting photo; the cutting photo is a high-definition digital image of a cleaned cutting sample taken by an industrial camera under a standard light source; The structured modal data includes logging-while-drilling data, well seismic trace data and element / mineral content data; the logging-while-drilling data includes GR, SP, RT, AC, DEN and CNL corresponding to the depth of the cutting sample; the well seismic trace data is three-dimensional seismic trace data near the depth point of the cutting sample; the element / mineral content data is obtained by XRF and XRD rapid analysis of the relative content of constant elements and minerals of the cutting sample.

4. The method of claim 1, wherein the method further comprises: The step S1 further comprises: S14, constructing a random lithology combination data set supporting prior constraints.

5. The method of claim 1, wherein: The step S2 is specifically: S21, using a general convolutional neural network as a backbone network, processing the input of the cutting photos, extracting the high-dimensional feature vector V of the unstructured data img ; all structured data is spliced into a one-dimensional vector, and the high-dimensional feature vector V of the structured data is extracted through a multi-layer perception network struct ; S22, through the unstructured data high-dimensional feature vector V img And the structured data high-dimensional feature vector V struct Calculate the correlation weight between the two modal features of unstructured data and structured data, and generate a fusion feature vector V that simultaneously fuses the two modal information fused ; S23, utilize the fusion feature vector V fused Output a probability distribution vector through a Softmax activation function; The expression of the probability distribution vector is: P=[p1,p2,...,p n ] where n is the total number of lithology classes, p i represents the probability that the sample belongs to the i-th lithology class, i = 1,..., n. S24, using the multi-modal data set constructed in step S1, taking the cross-entropy loss function as the objective function, and optimizing and training the parameters of the whole model through the back propagation algorithm.

6. The method of claim 5, wherein the prior constraints and expert interaction multi-modal intelligent LWD lithology identification method is characterized by: The unstructured high-dimensional feature vector V img Includes color, texture, grain morphology information of the cuttings photo.

7. The method of claim 5, wherein the priori constraints and expert interaction multi-modal intelligent LWD lithology identification method is characterized in that: The high-dimensional feature vector V of the unstructured data img and the high-dimensional feature vector V of the structured data struct The method for realizing multi-modal data fusion includes: using a cross-modal attention mechanism, using feature splicing and directly sending into an MLP, or using a multi-head attention mechanism for more fine-grained feature interaction.

8. The method of claim 1, wherein the method further comprises: The constraint list includes prior knowledge constraints and expert knowledge constraints; the prior knowledge constraints are geological rules.

9. The method of claim 1, wherein: The establishment method of the constraint list is: using a large language model to extract the input text description natural language to generate a possible lithology combination list and complete the establishment of the constraint list.

10. The method of claim 1, wherein: The final output probability P final The calculation formula is: P final = Softmax (P O M).

Images