Curve discrete self-supervised representation learning method for well logging cross-domain migration interpretation

Abstract

The present application relates to the field of oil and gas exploration geophysical well logging intelligent interpretation, and discloses a kind of discrete self-supervised representation learning method for logging cross-domain migration interpretation, which configures exclusive encoder and codebook for each type of logging curve, and discretizes continuous waveform slice into fine-grained token and quantization vector;Quantization vector is inverted into prototype waveform by decoder, macro token is generated by clustering fine-grained token based on waveform similarity, and global vocabulary is constructed for each type of logging;Self-supervised pre-training is carried out based on teacher-student dual branch architecture, and student network is optimized;After pre-training, the downstream task migration is carried out in the mode of main trunk freezing or parallel adapter fine-tuning, and the prediction result is output by deep dimension multi-curve dynamic weighted fusion.The present application can effectively distinguish the physical mechanism of different logging curves, adapt to the scene of label scarcity, curve loss and cross-region data distribution offset, and improve the cross-domain generalization ability of logging interpretation model.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation technology of geophysical logging in oil and gas exploration, specifically to a discrete self-supervised representation learning method for cross-domain migration interpretation of logging curves. It is applicable to reservoir evaluation tasks such as lithology classification and porosity prediction in scenarios involving multiple basins, multiple acquisition ages, scarce labels, missing logging curves, and geological regional distribution shifts. Background Technology

[0002] Well logging data is a core data source for formation property evaluation and reservoir characterization. Decades of exploration have accumulated a massive amount of unlabeled, multi-curve well logging archives, covering different basins and acquisition dates. However, this historical well logging data generally faces three major industry pain points: First, core labeling is costly, and many old and remote exploration wells lack effective labels; second, due to factors such as instrument malfunctions and differences in acquisition methods, some well logging curves are missing to varying degrees; third, reservoir properties differ significantly between different geological regions, resulting in numerous data distribution offsets, making it difficult for models trained in a single work area to generalize across regions.

[0003] Existing machine learning-based well logging interpretation methods are mainly divided into two categories: supervised deep learning and general time-series basic models. Supervised deep learning methods rely on massive amounts of labeled data for training, and the model performance is highly coupled with the data distribution of a specific work area. Once faced with a new basin or a new batch of acquisitions, the interpretation accuracy will decrease, and it cannot effectively reuse massive amounts of unlabeled historical well logging archives. General time-series basic models treat well logging curves with different physical measurement mechanisms, such as natural gamma (GR), sonic transit time (AC), volumetric density (DEN), neutron porosity (CNL), and deep resistivity (RT), as equivalent channels for unified encoding. This ignores the essential differences between various well logging methods in terms of measurement principles, dimensions, and dynamic response ranges, causing well logging signals with similar forms but completely different physical meanings to be mapped to the same latent space, leading to physical semantic confusion. In complex working conditions such as missing curves and regional offsets, the generalization ability is seriously insufficient.

[0004] Furthermore, existing methods based on discrete quantization representation (such as VQ-VAE and its variants) use a globally shared codebook to uniformly encode all logging curves, making it difficult to distinguish the physical interpretation characteristics of different logging types. At the same time, these methods lack structured pre-training constraints oriented towards the characteristics of logging data, resulting in limited model transferability in cross-work area and curve-missing scenarios.

[0005] Therefore, there is an urgent need for a well logging characterization learning method that can distinguish the physical mechanisms of different logging curves, adapt to scenarios with scarce labels and missing curves, and improve the ability to migrate across geological regions. Summary of the Invention

[0006] This invention addresses the technical problems of existing well logging interpretation models, such as scarce labels, missing well logging curves, data distribution shifts across geological regions, and the inability of general time series models and discrete quantization coding methods to distinguish different well logging physical mechanisms. It provides a discrete self-supervised representation learning method for well logging cross-domain transfer interpretation.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A discrete self-supervised representation learning method for well logging cross-domain migration interpretation includes: Each type of logging curve is configured with a dedicated encoder and codebook. The original waveform slices of each type of logging curve are mapped into hidden vectors by the encoder. By quantizing their respective codebooks, fine-grained tokens and corresponding quantization vectors are obtained. The quantization vectors corresponding to the fine-grained tokens in various logging curves are inverted into prototype waveforms by a dedicated decoder. Based on the similarity between prototype waveforms, the fine-grained tokens of the same type of logging curves are clustered and merged to generate macro tokens. A segmented and isolated global vocabulary that does not overlap is constructed for various logging curves. Self-supervised pre-training is performed based on a teacher-student dual-branch architecture. The teacher inputs a complete sequence, and the student inputs a sequence after structured masking, random curve discarding, and channel order perturbation. Macro tokens at corresponding positions in the global vocabulary are used as prediction supervision labels. The student network is optimized by combining mask token prediction, waveform reconstruction, local representation alignment, and global representation alignment loss. After completing the pre-training of the student network, the downstream task migration is carried out using either backbone freezing or parallel adapter fine-tuning mode, and the prediction results are output through dynamic weighted fusion of multiple curves in the depth dimension.

[0008] In one embodiment, configuring a dedicated encoder and codebook for each type of logging curve specifically includes: Encoder for well logging curves A dual-branch structure of amplitude and morphology is adopted: the amplitude branch extracts the mean and standard deviation of the original waveform slices of various logging curves and generates amplitude descriptors through a multilayer perceptron; the morphology branch processes the original waveform slices of various logging curves... After self-normalization, local morphological features are extracted via 1D convolution. The concatenated features from the two branches are then projected onto a latent vector. .

[0009] In one embodiment, obtaining fine-grained tokens and corresponding quantization vectors by quantizing each codebook specifically includes: Nearest neighbor quantization is used in the normalized vector space: ; ; in, As an implicit vector, For the dedicated codebook of the c-th type well logging curve, the first... codebook vectors It is the F2 norm. For the single-curve codebook capacity, To obtain fine-grained tokens through quantization, This is the quantization vector.

[0010] In one embodiment, the step of inverting the quantization vectors corresponding to fine-grained tokens in various logging curves into prototype waveforms via a dedicated decoder, and clustering and merging fine-grained tokens of similar logging curves based on the similarity between prototype waveforms to generate macro tokens specifically includes: No. Well logging curve A set of fine-grained token indices is ,in, and This represents any two fine-grained token indices in the set of fine-grained token indices, where the fine-grained token indexes are... and The corresponding codebook vectors are respectively input to the first A dedicated decoder for logging curves is used to obtain the corresponding prototype waveform; using the first... The prototype waveform explanation for each fine-grained token. When the prototype waveform corresponds to a fine-grained token, the corresponding directional determination coefficient is... for: ; in, Indicates the first In the type of well logging curve, the first The prototype waveform corresponding to each fine-grained token is in the... Waveform values ​​at each sampling location; Indicates the first In the type of well logging curve, the first The prototype waveform corresponding to each fine-grained token is in the... Waveform values ​​at each sampling location; Indicates the first The average value of the prototype waveform corresponding to each fine-grained token: ; in, Indicates the sampling position index in the prototype waveform; This indicates the number of sampling points contained in the original waveform slice and its corresponding prototype waveform. The distance is obtained after symmetry. ; Employing fully connected agglomerative clustering, when the... Fine-grained tokens and the first Distance of fine-grained tokens At that time, the two fine-grained tokens are merged to finally generate a mapping relationship. ; in, Indicates the first Mapping relationship between fine-grained tokens and macro tokens in well logging curves; Indicates the first The total number of fine-grained tokens for well logging-like curves; Indicates the first The total number of macro tokens obtained after clustering and merging well logging curves.

[0011] In one embodiment, the construction of a segmented, isolated global vocabulary for various well logging curves that does not overlap specifically includes: According to the preset order of logging curve types, a global offset is set for each type of logging curve. The global offset of a well logging curve is represented as: ; in, Indicates the first The initial offset of the well logging curve in the global vocabulary; Indicates the sorting position is at the 1st position. The first before the type of logging curve The total number of macro tokens for well logging-like curves; No. Local macro token index of well logging curve Converted to a global macro token index using the following mapping: ; in, Indicates the first Local macro token index in well logging curves The corresponding global macro token index; Indicates the first Local macro token index within the well logging curve vocabulary; Indicates the first The total number of macro tokens for well logging-like curves; No. The vocabulary interval corresponding to the well logging curve in the global vocabulary is represented as follows: ; The global prediction vocabulary is represented as follows: ; The size of the global prediction vocabulary is expressed as: ; in, Indicates the first Macro token word range corresponding to well logging curves; This represents a global prediction vocabulary formed by splicing macro token vocabulary intervals from various well logging curves; Represents a set of logging curve types; Indicates the index of well logging curve type; This indicates the total number of macro tokens in the global prediction vocabulary; various logging curves occupy non-overlapping continuous intervals in the global prediction vocabulary.

[0012] In one embodiment, in the teacher-student dual-branch architecture, the teacher network weight... Student network weight The update is performed using an exponential moving average, with the following update rules: ; in, The coefficient is the exponential moving average decay coefficient, and the teacher network does not participate in backpropagation gradient updates.

[0013] In one embodiment, the joint mask token prediction, waveform reconstruction, local representation alignment, and global representation alignment loss optimize the student network, specifically including: The masked macro token prediction loss is calculated only at the mask position in the student branch: ; in, This indicates the loss prediction using a masked macro token. This represents the set of mask positions in the student branch input sequence; Indicates the number of mask positions; Indicates the mask position index; Indicates the student network location The predicted logical value vector output for the global prediction vocabulary; Indicates the student network location True global macro token tag The output predicted logic value; Indicates the student network location The first item in the global vocabulary The predicted logical value output by each macro token; This represents the macro token index in the global prediction vocabulary; This represents a global prediction vocabulary formed by splicing macro token vocabulary intervals from various well logging curves; Indicates the mask position Corresponding real global macro token tag: ; in, Indicates position The original waveform slice is processed by the first Fine-grained token index obtained after quantization of the codebook specific to well logging curves; Indicates the first Mapping relationship between fine-grained tokens and macro tokens in well logging curves; Indicates the first Mapping relationship between local macro token indexes and global macro token indexes for well logging curves; Waveform reconstruction loss is decomposed into morphological loss based on relative fluctuation patterns. and amplitude reference loss based on overall reference constraints Weighted combination: ; This is the balance coefficient; The local representation alignment loss is expressed as: ; in, This represents the local alignment loss; This represents the set of mask positions in the student branch; Indicates the number of mask positions; This represents the mask position index within the student branch; Indicates the student network location The hidden state of the output; Indicates the student branch position Index mapping between positions of the same logging curve type and the same depth location in the teacher branch; Indicates the location of the teacher network The hidden state of the output; and The projection heads represent the local representations of the student network and the teacher network, respectively. Indicates stopping the gradient operator; This represents the function for calculating cosine similarity. The global representations of the student network and the teacher network are respectively expressed as: ; ; The global representation alignment loss is expressed as: ; in, This represents the global alignment loss; and Let each represent the set of valid non-missing positions in the student branch and the teacher branch, respectively. This indicates that the hidden states corresponding to valid non-missing locations are averaged. and These represent the global representations of the student network and the teacher network obtained by average pooling, respectively. and The global representation projection heads represent the student network and the teacher network, respectively; Indicates stopping the gradient operator; This represents the function for calculating cosine similarity.

[0014] In one embodiment, morphological loss for: ; Amplitude reference loss for: ; These represent the original waveform slice and the reconstructed slice value, respectively. , For decoder, These are the mean values ​​of the original waveform slices and the mean values ​​of the reconstructed slices, respectively. The length of the original waveform slice.

[0015] In one embodiment, the downstream task migration using backbone freezing or parallel adapter fine-tuning mode specifically includes: In the parallel adapter fine-tuning mode, bottleneck adapters are set as bypasses in each Transformer block of the pre-trained student network. The backbone parameters of the pre-trained Transformer blocks are kept frozen, and the backbone output and the bottleneck adapter output are superimposed to obtain the adapted block output. ; The output of the bottleneck adapter is represented as follows: ; in, This represents the input representation of the current Transformer block; This represents the output representation of the Transformer block after adjustment by the parallel adapter; This indicates a pre-trained Transformer block whose backbone parameters remain frozen; This represents a bottleneck adapter configured in parallel with a pre-trained Transformer block. Presentation layer normalization operation; This represents the dimension reduction projection matrix that maps the input representation from the backbone hidden dimension to the bottleneck dimension. Represents the activation function of the Gaussian error linear unit; This represents the up-dimensional projection matrix that maps the bottleneck dimension representation back to the hidden dimension of the backbone. This indicates a random deactivation operation performed on the adapter branch output; This represents the learnable scaling factor that controls the output injection strength of the adapter.

[0016] Compared with the prior art, the beneficial technical effects of the present invention are: 1. This invention adopts a curve-independent discrete coding method to configure a dedicated codebook for different types of well logging, which can effectively distinguish the measurement physical mechanism of various types of well logging and avoid semantic confusion between different well logging signals; at the same time, the encoder-decoder structure can realize well logging waveform reconstruction, making the discrete representation geophysical interpretable, which is convenient for field geologists to understand and apply.

[0017] 2. This invention sets up a segmented and isolated global vocabulary structure, combined with curve type embedding, to constrain the prediction range of the model, reduce the probability of erroneous predictions across well logging types, and improve the discriminativeness and stability of the representation space.

[0018] 3. This invention introduces multiple data perturbation methods, including structured masking, random well logging discarding, and channel order perturbation. It also combines a teacher-student dual-branch architecture with multiple joint loss constraints to conduct self-supervised pre-training, simulating various abnormal data conditions in the field. This enhances the model's adaptability to scenarios with missing curves and regional distribution shifts, and improves the overall performance of the model under label scarcity conditions.

[0019] 4. This invention designs two downstream model fine-tuning modes to adapt to different field application scenarios: for scenarios with large-area missing logging curves, the main trunk freezing mode is used to ensure the stability of model operation; for scenarios with complete curve data, the parallel adapter fine-tuning mode is used to utilize pre-trained representation capabilities to effectively improve lithology identification accuracy and reduce porosity prediction errors, meeting the needs of reservoir evaluation in different work areas. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of the present invention.

[0021] Figure 2 This is a schematic diagram of the independent discrete encoding structure of the curve in this invention; the diagram illustrates the process of single well logging encoding, quantization, decoding, token clustering, and vocabulary construction.

[0022] Figure 3 This is a schematic diagram of the framework for self-supervised pre-training of the teacher-student architecture in this invention; the diagram illustrates the dual-branch structure, data perturbation, parameter update, and multi-loss constraint logic.

[0023] Figure 4This is a schematic diagram of the downstream dual-mode fine-tuning and multi-curve fusion prediction structure in this invention. The diagram illustrates the transfer method of the pre-trained backbone model in the downstream task of the target work area. A parallel adapter fine-tuning mode can be adopted to improve the scene adaptability while maintaining the stability of the backbone parameters. Finally, the final prediction result is output through dynamic weighted fusion of multi-curves in the depth dimension.

[0024] Figure 5 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] like Figure 5 As shown, a discrete self-supervised representation learning method for well logging cross-domain transfer interpretation includes the following steps: S1 configures a dedicated encoder and codebook for each type of logging curve. The original waveform slices of each type of logging curve are mapped into hidden vectors by the encoder. By quantizing their respective codebooks, fine-grained tokens and corresponding quantization vectors are obtained. S2, the quantization vectors corresponding to the fine-grained tokens in various logging curves are inverted into prototype waveforms by a dedicated decoder. Based on the similarity between prototype waveforms, the fine-grained tokens of the same type of logging curves are clustered and merged to generate macro tokens, and a segmented isolated global vocabulary that does not overlap is constructed for various logging curves. S3 is based on a teacher-student dual-branch architecture for self-supervised pre-training. The teacher inputs a complete sequence, and the student inputs a sequence after structured masking, random curve discarding, and channel order perturbation. The macro tokens at the corresponding positions in the global vocabulary are used as prediction supervision labels. The student network is optimized by combining mask token prediction, waveform reconstruction, local representation alignment, and global representation alignment loss. S4, after completing the pre-training of the student network, uses backbone freezing or parallel adapter fine-tuning mode to migrate downstream tasks, and outputs the prediction results through deep-dimensional multi-curve dynamic weighted fusion.

[0027] The present invention will be described in detail below in several parts.

[0028] 1. Perform unified preprocessing on multi-source logging data.

[0029] (1) Data source division: The dataset is rigorously split into an unlabeled pre-training set and a labeled downstream evaluation set. The well locations in these two datasets are completely separate to avoid data leakage and ensure the stability of the model's transfer performance. The pre-training set integrates publicly available logging resources from multiple regions, selecting well data containing at least four valid logging curves to form a sufficient number of effective training samples. The downstream evaluation set uses independently labeled datasets from different geological blocks and data sources, used for performance validation in lithology classification and porosity prediction tasks, respectively. This invention standardizes seven mainstream logging types for oil and gas exploration: natural gamma (GR), spontaneous potential (SP), acoustic transit time (AC), deep resistivity (RT), bulk density (DEN), neutron porosity (CNL), and well diameter (CAL).

[0030] (2) Data cleaning and standardization: Null values ​​and invalid instrument codes were removed, and outliers were filtered based on value range and percentile; all well logs were resampled to vertical resolution. Missing intervals are retained without interpolation; the resistivity RT is logarithmically transformed to eliminate the influence of orders of magnitude span.

[0031] Block Z-score normalization and P99 percentile scaling: For well logging data of the same type within a single block, block standardization is performed first. ; , For blocks Inner curve The data mean and standard deviation are calculated; then, a second scaling is performed based on the P99 quantile of the curve in this block to suppress extreme value interference; downstream tasks only use training well statistics to complete normalization, eliminating the leakage of test set information.

[0032] 2. Different types of logging curves are independently and discretely encoded to construct a dedicated codebook for each curve.

[0033] like Figure 2 As shown, this step configures an independent VQ-VAE segmentation branch for each type of well logging, with different wells having their own dedicated codebook, thus isolating the hidden space of well logging with different measurement mechanisms. The coding process for a single well logging entry is as follows: Assume available well logging sets For any logging type Configure independent encoder decoder Exclusive codebook , This represents the capacity of a single-curve fine-grained codebook.

[0034] (1) Input length Depth window logging slices encoder Employing a dual-branch structure based on amplitude and morphology: Amplitude branch: Extract slice mean and standard deviation and generate amplitude descriptors using a lightweight MLP; Morphological branching: Slices are self-normalized to eliminate amplitude interference, and local morphological features are extracted through 1D convolution; the two feature branches are concatenated and projected to obtain continuous latent vectors. .

[0035] (2) The discrete token index and quantization vector are obtained by normalizing the vector space nearest neighbor quantization: , ; For fine-grained token indexing, This is the codebook quantization vector.

[0036] (3) Decoder reconstructs the original logging slices: ; The decoder adopts a residual MLP structure, which can back-derive the original logging waveform from any codebook vector, ensuring that the token has an interpretable physical form.

[0037] (4) Single-curve VQ-VAE joint loss function: ; For slices Reconstruction loss measures the point-by-point difference between the curve reconstructed by the Decoder and the original curve; To achieve vector quantization loss, the encoder output z is actively moved closer to the selected codebook vector e, while the codebook vector also moves closer to z. The codebook is updated using EMA exponential moving average. Low-usage codebook vectors are soft-refreshed online to improve codebook utilization. Codebook vectors that have not been selected for a long time are reset to the vicinity of the active region of the current encoder output, so that the codebook capacity is fully utilized.

[0038] 3. Perform macro-token clustering based on discrete encoding results and construct a segmented and isolated global vocabulary.

[0039] Based on the various output fine-grained discrete tokens from well logging, further macro-token clustering and global vocabulary construction are carried out. Single-curve fine-grained tokens suffer from redundancy and sparse, long-tailed distribution, which is detrimental to stable supervised learning during subsequent sequence-level pre-training. Therefore, this step clusters and merges fine-grained tokens based on the similarity of well logging decoding waveforms to generate macro-tokens with similar morphological features, and constructs a segmented and isolated global vocabulary. This provides standardized, highly stable discrete supervision labels and a unified prediction space for subsequent structure-aware self-supervised pre-training.

[0040] (1) Macro token clustering based on signal domain waveform similarity well logging All fine-grained indexes The decoder inverts to obtain the original waveform: .

[0041] Two-way decision coefficient Characterizing the waveform similarity between two token prototypes: .

[0042] Symmetricization yields the distance: , .

[0043] Fully connected agglomerative clustering is used, with a distance threshold d = 0.4 (i.e., similarity R² > 0.6). When the distance d between two tokens is less than 0.4, they are merged. The final mapping relationship is generated as follows: ; For curves Macro token total count, fine-grained token pass Mapped to the corresponding macro token, serving as a pre-trained discrete supervision label.

[0044] (2) Segmented isolation of global vocabulary construction: Set global offset according to logging type , For the sum of all logging macro tokens beforehand, single curve Macro token local index Map global indexes: Each curve occupies a non-overlapping interval in the global vocabulary: ; The global vocabulary is isolated from each logging domain, and the natural constraint tokens cannot be mismatched across curve types during single-class head prediction.

[0045] 4. Conduct structure-aware self-supervised pre-training based on a teacher-student framework to obtain general logging representations.

[0046] like Figure 3As shown, this step relies on the multi-curve logging data preprocessed in the first step and the segmented isolated global macro-token vocabulary constructed in the third step to conduct structure-aware self-supervised pre-training. A teacher-student dual-branch pre-training framework is adopted. The teacher branch inputs a complete standardized logging sequence, while the student branch inputs a logging sequence processed with structured masking, random logging discarding, and channel order perturbation. Discrete macro-token labels and multiple joint losses are combined to complete model pre-training. The model represents the depth correlation and cross-curve correlation of different logging curves through curve type encoding, depth position encoding, and rotation position encoding, thereby improving the adaptability of the obtained representation to scenarios with missing curves and regional distribution shifts.

[0047] (1) Input perturbation strategy: Structured masking: continuous depth masking within a single curve; synchronous masking across curves of the same depth, with the maximum mask ratio progressively increased to 0.5 during training. At the masked location, the curve value is replaced with a shared mask token, but the curve type embedding and depth position embedding are preserved.

[0048] Modal discarding: Some well logging types are randomly removed from the student end, simulating missing measured curves; Channel replacement: The logging order is fixed on the teacher's end, while the logging order is randomly shuffled on the student's end. The logging identity is distinguished by explicit curve embedding, thus breaking the dependence on the input position order.

[0049] (2) EMA teacher parameter update rules: Teacher network weights are updated by the moving average of student network exponents, with no backpropagation gradient. ; Teacher parameters Student parameters This is the EMA attenuation coefficient.

[0050] (3) Four pre-training loss functions (total loss) ): Masked macro token predicts loss : Cross-entropy is calculated only at the mask position. The mask location is the real global macro token label. For the set of mask locations: .

[0051] Mask slice reconstruction loss : Split-form loss Amplitude reference loss Weighted combination: ; ; ; These are the original and reconstructed slice values, respectively. These are the mean values ​​of the original and reconstructed slices, respectively. P is the balance coefficient, and P is the depth window length (number of sampling points) of the logging slice.

[0052] Local teacher-student representation alignment loss : Masking positions align student representations with homogeneous, same-depth teacher representations. Match teacher indexes to student locations. To stop the gradient operator: ; Where i is the index of a position in the student sequence, and ρ(i) is the corresponding index in the teacher sequence. Let i be the hidden state vector of the student Transformer at position i. Let be the hidden state vector of the teacher's Transformer at position ρ(i), and let cos denote the cosine similarity calculation. , An optional linear projection layer is used to map the hidden representations of students and teachers to the same contrast space.

[0053] Global sequence characterization alignment loss : Effective positions are aligned after global pooling. The sets of valid non-missing locations for students and teachers are respectively: ; ; .

[0054] (4) Overall pre-training objective function ; These are the loss weight coefficients for local alignment, reconstruction, and global alignment, respectively.

[0055] 5. Downstream dual-mode adaptive fine-tuning is performed based on the pre-trained backbone model, and lithology classification and porosity prediction are completed through multi-curve weighted fusion.

[0056] like Figure 4As shown, this step builds upon the backbone model pre-trained in step four, which possesses cross-domain robustness and physical semantic consistency, to construct a downstream dual-mode adaptive transfer learning framework adapted to different engineering application scenarios. Two schemes are set up: a fine-tuning mode with the backbone model fully frozen, and a fine-tuning mode with the backbone model frozen and paired with a parallel adapter. The former only trains the downstream task output head. Combining the data standardization rules and solidified physical representation characteristics of the curves fixed in step one, the pre-trained general representation is fine-tuned according to specific scenarios. Finally, the actual downstream task (such as lithology classification and porosity regression downstream prediction tasks) is completed through multi-curve adaptive weighted fusion.

[0057] (1) Parallel adapter structure: Transformer block parallel embedding bottleneck adapter, raw trunk output With adapter output Overlay: ; in, ; Input x first passes through Layer normalization, then through Dimensionality reduction projection, followed by GELU nonlinear activation and then... Dimensionality increase. By using Drop random deactivation, the adapter is forced to learn more robust features, preventing overfitting. This is a learnable scaling factor. The adapter is initialized to zero, and during the fine-tuning phase, it gradually deviates from the pre-trained weights.

[0058] (2) Depth-dimensional adaptive multi-curve weighted fusion prediction: Let the first The set of logging curve types available at each depth location is: For any available logging curve type The logging curve was on the 1st Each depth location corresponds to a representation input to the task output head, which independently generates a single curve prediction result: ; in, Indicates the depth location index; Indicates the index of well logging curve type; Indicates the first The set of actually usable logging curve types at each depth location; Indicates the first The logging curve of type 1 in the first The logging representation obtained after processing by a pre-trained backbone model and optional parallel adapter at a depth location; Indicates the output header of the downstream task; Indicates the first The logging curve of type 1 in the first Single-curve prediction results generated independently at each depth location. In lithological classification tasks, This represents the predicted logical value vector corresponding to each lithology category; in the porosity prediction task... This is a predicted value for continuous porosity; Each available logging curve generates prediction results independently. Missing or invalid logging curves are not included in the generation of prediction results for the current depth location or in subsequent fusion.

[0059] For the Any available logging curve type at any depth location The well logging characterization is input into a lightweight fusion weight generation network to obtain the corresponding unnormalized fusion weights: ; in, Indicates the first The logging curve of type 1 in the first Unnormalized fusion weights at each depth location; Indicates the first The logging curve of type 1 in the first The logging representation obtained after processing by a pre-trained backbone model and optional parallel adapter at a depth location; This represents a lightweight multilayer perceptron used to generate curve fusion weights; This represents the Sigmoid activation function, used to restrict the fusion weights to non-negative values.

[0060] The unnormalized fusion weights of all actually usable logging curves at the same depth location are normalized to obtain the normalized fusion weights: ; in, Indicates the first The logging curve of type 1 in the first Normalized fusion weights at each depth location; Indicates the first The curve index from the set of logging curve types is available at each depth location; Indicates the first The set of actually usable logging curve types at each depth location; This represents a numerically stable term, used to prevent the denominator from being zero in normalization calculations.

[0061] Based on the normalized fusion weights, the prediction results generated independently from each available logging curve at the same depth location are weighted and fused to obtain the... Final prediction results at each depth location: ; in, Indicates the first The final fusion prediction results at each depth location; Indicates the first The logging curve of type 1 in the first Normalized fusion weights at each depth location; Indicates the first The logging curve of type 1 in the first Single-curve prediction results generated independently at each depth location. For lithological classification tasks, This represents the fused lithology category prediction logical value vector; for the porosity prediction task, This represents the predicted value of continuous porosity after fusion.

[0062] For actual downstream tasks, in the lithology classification task, the fused depth point prediction results are used as the classification output, and the cross-entropy loss function is used to optimize the model; in the porosity prediction task, the fused continuous prediction values ​​are used as the regression output, and the regression loss is calculated only at effective depth points with core measured porosity labels.

[0063] Thus, this step, through dual-mode adaptive transfer and depth-dimensional multi-curve fusion prediction, enables the application of pre-trained general representations in specific reservoir evaluation tasks, allowing the model to improve the stability and applicability of interpretation results even under conditions of missing logging curves, regional distribution shifts, and limited labeled samples.

[0064] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0065] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0066] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0067] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0068] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A discrete self-supervised representation learning method for well logging cross-domain transfer interpretation, characterized in that, include: Each type of logging curve is configured with a dedicated encoder and codebook. The original waveform slices of each type of logging curve are mapped into hidden vectors by the encoder. By quantizing their respective codebooks, fine-grained tokens and corresponding quantization vectors are obtained. The quantization vectors corresponding to the fine-grained tokens in various logging curves are inverted into prototype waveforms by a dedicated decoder. Based on the similarity between prototype waveforms, the fine-grained tokens of the same type of logging curves are clustered and merged to generate macro tokens. A segmented and isolated global vocabulary that does not overlap is constructed for various logging curves. Self-supervised pre-training is performed based on a teacher-student dual-branch architecture. The teacher inputs a complete sequence, and the student inputs a sequence after structured masking, random curve discarding, and channel order perturbation. Macro tokens at corresponding positions in the global vocabulary are used as prediction supervision labels. The student network is optimized by combining mask token prediction, waveform reconstruction, local representation alignment, and global representation alignment loss. After completing the pre-training of the student network, the downstream task migration is carried out using either backbone freezing or parallel adapter fine-tuning mode, and the prediction results are output through dynamic weighted fusion of multiple curves in the depth dimension.

2. The discrete self-supervised representation learning method for well logging cross-domain migration interpretation according to claim 1, characterized in that, The process of configuring a dedicated encoder and codebook for each type of well logging curve specifically includes: Encoder for well logging curves A dual-branch structure of amplitude and morphology is adopted: the amplitude branch extracts the mean and standard deviation of the original waveform slices of various logging curves and generates amplitude descriptors through a multilayer perceptron; the morphology branch processes the original waveform slices of various logging curves... After self-normalization, local morphological features are extracted via 1D convolution. The concatenated features from the two branches are then projected onto a latent vector. .

3. The discrete self-supervised representation learning method for well logging cross-domain migration interpretation according to claim 1, characterized in that, The process of obtaining fine-grained tokens and corresponding quantization vectors by quantizing their respective codebooks specifically includes: Nearest neighbor quantization is used in the normalized vector space: ； ； in, As an implicit vector, For the dedicated codebook of the c-th type well logging curve, the first... codebook vectors It is the F2 norm. For the single-curve codebook capacity, To obtain fine-grained tokens through quantization, This is the quantization vector.

4. The discrete self-supervised representation learning method for well logging cross-domain transfer interpretation according to claim 1, characterized in that, The process involves inverting the quantization vectors corresponding to fine-grained tokens in various logging curves into prototype waveforms using a dedicated decoder, and then clustering and merging fine-grained tokens from similar logging curves based on the similarity between prototype waveforms to generate macro tokens. Specifically, this includes: No. Well logging curve A set of fine-grained token indices is ,in, and This represents any two fine-grained token indices in the set of fine-grained token indices, where the fine-grained token indexes are... and The corresponding codebook vectors are respectively input to the first A dedicated decoder for logging curves is used to obtain the corresponding prototype waveform; using the first... The prototype waveform explanation for each fine-grained token. When the prototype waveform corresponds to a fine-grained token, the corresponding directional determination coefficient is... for: ； in, Indicates the first In the type of well logging curve, the first The prototype waveform corresponding to each fine-grained token is in the... Waveform values ​​at each sampling location; Indicates the first In the type of well logging curve, the first The prototype waveform corresponding to each fine-grained token is in the... Waveform values ​​at each sampling location; Indicates the first The average value of the prototype waveform corresponding to each fine-grained token: ； in, Indicates the sampling position index in the prototype waveform; This indicates the number of sampling points contained in the original waveform slice and its corresponding prototype waveform. The distance is obtained after symmetry. ; Employing fully connected agglomerative clustering, when the... The fine-grained token and the first Distance of fine-grained tokens At that time, the two fine-grained tokens are merged to finally generate a mapping relationship. ; in, Indicates the first Mapping relationship between fine-grained tokens and macro tokens in well logging curves; Indicates the first The total number of fine-grained tokens for well logging-like curves; Indicates the first The total number of macro tokens obtained after clustering and merging well logging curves.

5. The discrete self-supervised representation learning method for well logging cross-domain migration interpretation according to claim 1, characterized in that, The construction of a segmented, isolated global vocabulary for various well logging curves, which does not overlap, specifically includes: According to the preset order of logging curve types, a global offset is set for each type of logging curve. The global offset of a well logging curve is represented as: ； in, Indicates the first The initial offset of the well logging curve in the global vocabulary; Indicates the sorting position is at the 1st position. The first before the type of logging curve The total number of macro tokens for well logging-like curves; No. Local macro token index of well logging curve Converted to a global macro token index via the following mapping: ； in, Indicates the first Local macro token index in well logging curves The corresponding global macro token index; Indicates the first Local macro token index within the well logging curve vocabulary; Indicates the first The total number of macro tokens for well logging-like curves; No. The vocabulary interval corresponding to the well logging curve in the global vocabulary is represented as follows: ； The global prediction vocabulary is represented as follows: ； The size of the global prediction vocabulary is expressed as: ； in, Indicates the first Macro token word range corresponding to well logging curves; This represents a global prediction vocabulary formed by splicing macro token vocabulary intervals from various well logging curves; Represents a set of logging curve types; Indicates the index of well logging curve type; This indicates the total number of macro tokens in the global prediction vocabulary; various logging curves occupy non-overlapping continuous intervals in the global prediction vocabulary.

6. The discrete self-supervised representation learning method for well logging cross-domain transfer interpretation according to claim 1, characterized in that, In the aforementioned teacher-student dual-branch architecture, the teacher's network weight Student network weight The update is performed using an exponential moving average, with the following update rules: ； in, The coefficient is the exponential moving average decay coefficient, and the teacher network does not participate in backpropagation gradient updates.

7. The discrete self-supervised representation learning method for well logging cross-domain transfer interpretation according to claim 1, characterized in that, The joint mask token prediction, waveform reconstruction, local representation alignment, and global representation alignment losses are used to optimize the student network, specifically including: The masked macro token prediction loss is calculated only at the mask position in the student branch: ； in, This indicates the loss prediction using a masked macro token. This represents the set of mask positions in the student branch input sequence; Indicates the number of mask positions; Indicates the mask position index; Indicates the student network location The vector of predicted logical values ​​output for the global prediction vocabulary; Indicates the student network location True global macro token tag The output predicted logic value; Indicates the student network location The first item in the global vocabulary The predicted logical value output by each macro token; This represents the macro token index in the global prediction vocabulary; This represents a global prediction vocabulary formed by splicing macro token vocabulary intervals from various well logging curves; Indicates the mask position Corresponding real global macro token tag: ； in, Indicates position The original waveform slice is processed by the first Fine-grained token index obtained after quantization of the codebook specific to well logging curves; Indicates the first Mapping relationship between fine-grained tokens and macro tokens in well logging curves; Indicates the first Mapping relationship between local macro token indexes and global macro token indexes for well logging curves; Waveform reconstruction loss is decomposed into morphological loss based on relative fluctuation patterns. and amplitude reference loss based on overall reference constraints Weighted combination: ； This is the balance coefficient; The local representation alignment loss is expressed as: ； in, This represents the local alignment loss; This represents the set of mask positions in the student branch; Indicates the number of mask positions; This represents the mask position index in the student branch; Indicates the student network location The hidden state of the output; Indicates the student branch position Index mapping between positions of the same logging curve type and the same depth location in the teacher branch; Indicates the location of the teacher network The hidden state of the output; and The projection heads represent the local representations of the student network and the teacher network, respectively. Indicates stopping the gradient operator; This represents the function for calculating cosine similarity. The global representations of the student network and the teacher network are respectively expressed as: ； ； The global representation alignment loss is expressed as: ； in, This represents the global alignment loss; and Let each represent the set of valid non-missing positions in the student branch and the teacher branch, respectively. This indicates that the hidden states corresponding to valid non-missing locations are averaged. and These represent the global representations of the student network and the teacher network obtained by average pooling, respectively. and The global representation projection heads represent the student network and the teacher network, respectively; Indicates stopping the gradient operator; This represents the function for calculating cosine similarity.

8. The discrete self-supervised representation learning method for well logging cross-domain migration interpretation according to claim 7, characterized in that, Shape loss for: ； Amplitude reference loss for: ； These represent the original waveform slice and the reconstructed slice value, respectively. , For decoder, These are the mean values ​​of the original waveform slices and the mean values ​​of the reconstructed slices, respectively. The length of the original waveform slice.

9. The discrete self-supervised representation learning method for well logging cross-domain migration interpretation according to claim 1, characterized in that, The downstream task migration using either backbone freezing or parallel adapter fine-tuning modes specifically includes: In the parallel adapter fine-tuning mode, bottleneck adapters are set as bypasses in each Transformer block of the pre-trained student network. The backbone parameters of the pre-trained Transformer blocks are kept frozen, and the backbone output and the bottleneck adapter output are superimposed to obtain the adapted block output. ； The output of the bottleneck adapter is represented as follows: ； in, This represents the input representation of the current Transformer block; This represents the output representation of the Transformer block after adjustment by the parallel adapter; This indicates a pre-trained Transformer block whose backbone parameters remain frozen; This represents a bottleneck adapter configured in parallel with a pre-trained Transformer block. Presentation layer normalization operation; This represents the dimension reduction projection matrix that maps the input representation from the backbone hidden dimension to the bottleneck dimension. Represents the activation function of the Gaussian error linear unit; This represents the up-dimensional projection matrix that maps the bottleneck dimension representation back to the hidden dimension of the backbone. This indicates a random deactivation operation performed on the adapter branch output; This represents the learnable scaling factor that controls the output injection strength of the adapter.

Images