Shale oil geochemical logging free hydrocarbon content analysis method and intelligent sensor

Abstract

The application discloses a shale oil geochemical logging free hydrocarbon content analysis method and an intelligent sensor, relates to the technical field of intelligent sensors, and comprises the following steps: acquiring geochemical logging original signal data and synchronously collecting element logging data; pre-processing the geochemical logging original signal data to obtain corrected signal data; determining a quantitative lithology factor, and inputting the corrected signal data and the quantitative lithology factor into a lithology self-adaptive correction model to obtain lithology-corrected equivalent hydrocarbon signal data; inputting the lithology-corrected equivalent hydrocarbon signal data into an inversion model to obtain free hydrocarbon content prediction data, and performing cross verification and output according to gas logging data of a target well section. The technical problem that the existing lithology correction model is rigid and multi-source data fusion is insufficient, thereby limiting the prediction accuracy of free hydrocarbon content, is solved.

Description

Technical Field

[0001] This application relates to the field of intelligent sensor technology, specifically to a method for analyzing free hydrocarbon content in shale oil geochemical logging and an intelligent sensor. Background Technology

[0002] As shale oil exploration and development progresses, the measurement accuracy of geochemical logging technology affects the reliability of resource assessment.

[0003] However, traditional geochemical logging analysis methods mainly rely on pyrolysis chromatography to quantitatively detect free hydrocarbons, pyrolysis hydrocarbons, and residual hydrocarbons in cuttings samples. They use fixed lithology correction coefficients or empirical formulas for correction, which are difficult to adapt to the frequent alternation of lithology in the vertical sequence. This leads to systematic biases in the calculation results of free hydrocarbon content and fails to capture the influence mechanism of dynamic lithology changes on the signal response. Summary of the Invention

[0004] This application provides a method for analyzing free hydrocarbon content in shale oil geochemical logging and an intelligent sensor, which solves the technical problem that the existing lithology correction model is rigid and the fusion of multi-source data is insufficient, resulting in limited accuracy in predicting free hydrocarbon content.

[0005] The technical solution to the above-mentioned technical problems in this application is as follows:

[0006] In a first aspect, this application provides a method for analyzing the free hydrocarbon content in shale oil geochemical logging, the method comprising: Obtain raw geochemical logging signal data of the target well section and simultaneously acquire elemental logging data, wherein the elemental logging data includes at least the content data of lithological indicator elements; The raw geochemical logging signal data is preprocessed to obtain corrected signal data; Based on the elemental logging data, a quantitative lithology factor is determined, and the corrected signal data and the quantitative lithology factor are input into a lithology adaptive correction model to obtain lithology-corrected equivalent hydrocarbon signal data. The lithology adaptive correction model is configured to dynamically select the correction path based on the lithological complexity. The lithology-corrected equivalent hydrocarbon signal data is input into the inversion model to obtain free hydrocarbon content prediction data, and cross-validation and output are performed based on the gas logging data of the target well section.

[0007] Secondly, this application provides a smart sensor for analyzing free hydrocarbon content in shale oil geochemical logging, comprising: The information acquisition module is used to acquire the original geochemical logging signal data of the target well section and simultaneously acquire elemental logging data, wherein the elemental logging data includes at least the content data of lithological indicator elements. The data correction module is used to preprocess the raw geochemical logging signal data to obtain corrected signal data; The model training module is used to determine the quantitative lithology factor based on the elemental logging data, and input the corrected signal data and the quantitative lithology factor into the lithology adaptive correction model to obtain the lithology-corrected equivalent hydrocarbon signal data. The lithology adaptive correction model is configured to dynamically select the correction path based on the lithological complexity. The result output module is used to input the lithology-corrected equivalent hydrocarbon signal data into the inversion model to obtain free hydrocarbon content prediction data, and to perform cross-validation and output based on the gas logging data of the target well section.

[0008] This application provides one or more technical solutions, which have at least the following technical effects or advantages: This application provides a method and intelligent sensor for analyzing free hydrocarbon content in shale oil through geochemical logging. First, by simultaneously acquiring raw geochemical logging signal data and elemental logging data, it achieves collaborative acquisition of multi-source geochemical information, overcoming the limitations of traditional methods that rely solely on pyrolysis chromatographic signals. Second, based on a lithology adaptive correction model constructed using quantitative lithology factors, it dynamically selects the correction path according to lithological complexity. For single-lithology regions, a lightweight end-side model is used for rapid response; for mixed-lithology regions, a hierarchical architecture combining cloud-based meta-learning and end-side incremental fine-tuning is employed, effectively solving the technical bottleneck of fixed correction coefficients being unable to adapt to frequent vertical lithological alternations. Third, by mining non-lithological indicator elements through a factor encoder, quantitative coding factors are extracted and integrated into the lithological characterization system, enhancing the model's ability to perceive potential geological features. Finally, gas logging data is introduced for dual cross-validation of trend consistency and fluid property logical consistency, constructing a multi-verification mechanism for the prediction results, improving the reliability and geological applicability of free hydrocarbon content prediction.

[0009] Through the above technical solutions, this application realizes a paradigm shift in shale oil geochemical logging free hydrocarbon content analysis from static empirical correction to dynamic intelligent sensing, improving the accuracy and stability of free hydrocarbon content prediction under complex lithological conditions, and providing technical support for shale oil resource evaluation. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1This is a schematic flowchart of a method for analyzing free hydrocarbon content in shale oil geochemical logging provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an intelligent sensor for analyzing free hydrocarbon content in shale oil geochemical logging, provided in an embodiment of this application.

[0012] The components represented by each number in the attached diagram are explained below: Information acquisition module 11, data correction module 12, model training module 13, and result output module 14. Detailed Implementation

[0013] This application provides a method for analyzing free hydrocarbon content in shale oil geochemical logging and an intelligent sensor to address the technical problem of limited accuracy in predicting free hydrocarbon content caused by rigidity in existing lithology correction models and insufficient fusion of multi-source data.

[0014] Example 1, as Figure 1 As shown in the embodiments of this application, a method for analyzing the free hydrocarbon content in shale oil geochemical logging is provided, including: S10: Acquire the raw geochemical logging signal data of the target well section and simultaneously collect elemental logging data, wherein the elemental logging data includes at least the content data of lithological indicator elements; In this embodiment of the application, continuous pyrolysis analysis of cuttings samples is performed using an on-site geochemical logging instrument to obtain raw geochemical logging signal data, including time-series data such as pyrolysis temperature sequence signals, hydrocarbon release rate curves, and baseline drift characteristics. The sampling interval is usually set to one analysis point per meter or every 0.5 meters.

[0015] Meanwhile, elemental logging data, including the content data of lithological indicator elements, such as the mass fractions of various elements such as silicon, aluminum, iron, calcium, magnesium, potassium, sodium, titanium, manganese, sulfur, barium, and strontium, were obtained through X-ray fluorescence spectroscopy analysis. Among them, the silicon-aluminum ratio, iron-magnesium combination, and calcium content fluctuations were pre-labeled as lithological indicator elements to characterize the geochemical fingerprints of different lithological units such as mudstone, shale, siltstone, and calcareous interlayers.

[0016] Specifically, the synchronous acquisition mechanism ensures strict spatial correspondence between the two types of data by aligning the unified depth index with the timestamp, thus eliminating depth misalignment caused by differences in sampling frequency.

[0017] S20: Preprocess the raw geochemical logging signal data to obtain corrected signal data; In this embodiment of the application, the raw geochemical logging signal data collected above is preprocessed. First, outliers in the raw geochemical logging signal data are removed. Then, the pyrolysis temperature sequence signal is baseline corrected to eliminate systematic errors introduced by instrument drift and sample matrix effect. Finally, the signal intensity of different sampling points is mapped to a unified dimension through depth normalization to obtain the corrected signal data.

[0018] Specifically, step S20 in the method includes: Combining a preset sliding window and the local outlier factor method, the original geochemical logging signal data is traversed to identify outliers, and the outliers are replaced with the median value under the sliding window. Adaptive baseline fitting is performed on the replaced original geochemical logging signal data, and the baseline fitting result is subtracted accordingly to obtain the corrected signal data.

[0019] In this embodiment, firstly, outlier identification is performed on the raw geochemical logging signal data by combining a preset sliding window and the local outlier factor method. Specifically, the width of the sliding window is dynamically set according to the average lithological layer thickness of the target well section, typically taking 3 to 5 times the number of sampling points corresponding to the thickness of the smallest lithological unit, to ensure that excessive smoothing is avoided while capturing local data features. The local outlier factor method identifies anomalous jump points caused by sample contamination, momentary instrument malfunction, or operational errors by calculating the degree of deviation of each data point from its neighborhood density. Furthermore, for data points determined to be anomalous, the median value of non-anomalous data within the sliding window is used for replacement to preserve the robustness of the data distribution.

[0020] Secondly, adaptive baseline fitting is performed on the updated geochemical logging raw signal data. The adaptive baseline fitting adopts a penalized least squares smoothing algorithm, which balances baseline tracking capability and noise suppression by adjusting the smoothing parameter.

[0021] Specifically, the low-frequency trend of the pyrolysis temperature sequence is used as a baseline reference. The fitting curve is iteratively optimized until the convergence condition is met. Then, the baseline fitting result is subtracted from the original signal to eliminate the baseline drift caused by differences in sample loading, instrument sensitivity decay and ambient temperature fluctuations. Finally, the corrected signal data with uniform dimensions is output.

[0022] S30: Based on the elemental logging data, determine the quantitative lithology factor, and input the corrected signal data and the quantitative lithology factor into the lithology adaptive correction model to obtain the lithology-corrected equivalent hydrocarbon signal data, wherein the lithology adaptive correction model is configured to dynamically select the correction path according to the lithological complexity; In this embodiment of the application, the elemental logging data obtained by S10 analysis is standardized to eliminate the dimensional influence caused by the difference in the content of different elements, and a quantitative lithology factor is obtained. The corrected signal data and the quantitative lithology factor are then input into the lithology adaptive correction model. Through lithology complexity assessment, the equivalent hydrocarbon signal data after lithology correction is obtained, and the optimal correction path is dynamically selected.

[0023] The determination of quantitative lithological factors based on the elemental logging data includes: The content data of lithological indicator elements are extracted from the elemental logging data and normalized to obtain the normalized quantitative value of each lithological indicator element. The normalized quantitative values ​​of each lithological indicator element are combined into multidimensional time-series data and determined as the quantitative lithological factor.

[0024] In this embodiment of the application, firstly, pre-labeled lithological indicator elements, including the content data of key elements such as silicon, aluminum, iron, magnesium, and calcium, are extracted from the elemental logging data. The Z-score standardization method is used to normalize the data, converting the content of each element into a dimensionless value with a mean of zero and a standard deviation of one, thereby obtaining the normalized quantitative value of each lithological indicator element.

[0025] For example, for the silicon-aluminum ratio, an index reflecting the intensity of clastic input is constructed by calculating the ratio of the normalized quantified values ​​of silicon and aluminum; for the iron-magnesium combination, a weighted linear combination of iron and magnesium is used to characterize the redox state of the sedimentary environment; and for the fluctuation of calcium content, the boundary characteristics of carbonate rock interlayers are captured by first-order difference operations.

[0026] Subsequently, the normalized quantitative values ​​of each lithological indicator element were arranged according to the depth sequence and combined into multidimensional time-series data, which were determined as quantitative lithological factors. These quantitative lithological factors not only retain the lithological discrimination information of a single element, but also enhance the distinguishability of lithological units through multidimensional combination.

[0027] Furthermore, determining the quantitative lithological factors based on the elemental logging data also includes: The multidimensional time-series data of the quantified values ​​corresponding to the non-lithological indicator elements in the element logging data are defined as quantitative potential factors. A factor encoder is constructed, comprising an autoencoder layer and a decoder layer. The autoencoder layer takes the quantitative latent factor as input and outputs a quantitative encoded factor. The decoder layer takes the quantitative encoded factor as input and outputs a quantitative reduced factor. Construct a joint loss function, and train the factor encoder based on the joint loss function and the sample element logging data. The joint loss function includes a reconstruction loss term based on the mean square error and a central level term based on the dimensional contribution rate. The quantified values ​​of the portion of the elemental logging data that is not the lithological indicator element are input into the autoencoder layer of the factor encoder, and the quantitative coding factor is added to the quantitative lithological factor.

[0028] In this embodiment of the application, firstly, the elements other than lithological indicator elements in the elemental logging data are defined, including trace elements such as titanium, manganese, sulfur, barium, strontium, vanadium, nickel, chromium, copper, and zinc. The multidimensional time series data of their corresponding quantitative values ​​are quantitative potential factors. Although other elements are not direct lithological indicator indicators, they contain potential information on the chemical properties of sedimentary media, organic matter abundance, and diagenetic evolution stages. Through depth feature extraction, they can be transformed into auxiliary factors that enhance the ability to characterize lithology.

[0029] Secondly, a factor encoder consisting of an autoencoder layer and a decoder layer is constructed. The autoencoder layer adopts a hybrid architecture of a one-dimensional convolutional neural network and a bidirectional long short-term memory network. It takes quantitative latent factors as input and extracts compressed deep features through multi-layer nonlinear transformations. The output dimension is one-third of the original input dimension of the quantitative encoded factors. The decoder layer adopts a symmetrical transposed convolution and inverse long short-term memory network structure. It takes quantitative encoded factors as input and reconstructs quantitative restored factors.

[0030] Furthermore, a joint loss function is constructed to guide the training process of the factor encoder. This joint loss function consists of two parts: a reconstruction loss term based on mean squared error, which measures the reconstruction error between the quantitative reconstructed factor and the original quantitative latent factor, ensuring the information fidelity of the encoding process; and a lumped level term based on dimensional contribution rate, which constrains the variance distribution of each dimension of the quantitative encoding factor, avoiding feature redundancy and promoting the saliency of key dimensions. Specifically, the lumped level term penalizes overly dispersed feature distributions by calculating the ratio of the variance of each dimension to the total variance, making the model tend to learn compact representations with clear geological significance.

[0031] Furthermore, based on the aforementioned joint loss function, the factor encoder is trained end-to-end using sample element logging data from multiple target well sections. The network parameters are optimized using a stochastic gradient descent algorithm until the loss function converges to a stable threshold. The trained factor encoder can then transform high-dimensional, noise-sensitive trace element data into low-dimensional, geologically semantically clear quantitative coding factors.

[0032] Finally, the quantified values ​​of the non-lithological indicator elements in the elemental logging data are input into the autoencoder layer of the trained factor encoder to obtain the output quantitative coding factor, and then spliced ​​into the aforementioned quantitative lithological factor constructed from lithological indicator elements to form an enhanced quantitative lithological factor that integrates explicit lithological indicators and implicit geological features.

[0033] The process includes constructing a joint loss function and training the factor encoder based on the joint loss function and sample element logging data, comprising: Define the mean square error between the quantitative latent factor and the quantitative reduction factor as the reconstruction loss term; Using the multiple coding dimensions of the quantitative coding factors as analysis dimensions, the principal component contribution rate obtained by principal component analysis of any batch of quantitative coding factor sets is defined as the contribution rate of the dimensions. Based on the dimensional contribution rate distribution formed by the multiple dimensional contribution rates, the concentration degree is calculated, and the reciprocal of the concentration degree is defined as the concentration level term; The joint loss function is obtained by weighting the reconstruction loss term and the central level term. Based on historical element logging data, sample element logging data is obtained, and the factor encoder is trained with the goal of minimizing the value of the joint loss function.

[0034] In this embodiment, firstly, the mean squared error between the quantitative latent factor and the quantitative reduction factor is defined as the reconstruction loss term, and the specific calculation formula is: Lr=(1 / N)× ||xi-x'i||², where xi represents the quantitative latent factor vector of the i-th sampling point, x'i represents the corresponding quantitative restoration factor vector, and N is the batch sample size. This loss term ensures that the encoder retains the key information of the original data while compressing the feature dimension.

[0035] Secondly, using multiple coding dimensions of the quantitative coding factors as analytical dimensions, principal component analysis is performed on any batch of quantitative coding factor sets to obtain the variance contribution rate of each principal component as the dimension contribution rate. Let the batch quantitative coding factor matrix be Z∈R^ (N×M) Where M is the number of encoding dimensions, and the feature value sequence λ1≥λ2≥...≥λM is obtained through singular value decomposition, then the contribution rate of the k-th dimension is ηk=λk / λi.

[0036] Secondly, the concentration is calculated based on the dimensional contribution rate distribution formed by multiple dimensions, and the Herfindahl-Hirschman Index is used to measure the degree of imbalance in the distribution, where the concentration H = ηk², the closer this value is to 1, the more concentrated the features are in a few dimensions; the closer it is to 1 / M, the more uniform the feature distribution. The concentration level term is defined as the reciprocal of the concentration degree, i.e., Ls = 1 / H. This design incentivizes the model to learn sparse yet discriminative feature representations, avoiding excessive dispersion of information across redundant dimensions.

[0037] Subsequently, a joint loss function is constructed by weighted fusion and reconstruction of the loss term and the lumped level term, namely L=α×Lr+β×Ls, where α and β are preset balance coefficients, which are usually initialized according to the magnitude ratio of the loss in the early stage of training and dynamically adjusted by an adaptive learning rate strategy during training to prevent a certain loss term from dominating the gradient direction.

[0038] Finally, a sample library was constructed based on historical elemental logging data, covering typical well sections with different sedimentary environments, maturity levels, and lithological combinations. Data augmentation techniques, including random depth migration, elemental content perturbation, and time stretching transformation, were used to expand sample diversity. With minimizing the joint loss function as the optimization objective, the Adam optimizer was used to train the factor encoder. An early stopping mechanism was implemented to monitor the validation set loss. Training was terminated and the model was reverted to the optimal parameter state if the validation loss did not decrease after ten consecutive training epochs, ultimately yielding a factor encoder model with good generalization ability.

[0039] Based on historical elemental logging data, sample elemental logging data is obtained, and the factor encoder is trained with the objective of minimizing the value of the joint loss function. Following this, the process further includes: Based on the completed training of the factor encoder, obtain the training validation record; Analyze the training and validation records to obtain the dimensional contribution rate distribution corresponding to the set level item; The dimensional contribution rate distribution is serialized, and combined with the preset cumulative contribution rate constraint, the first N dimensions whose sum of dimensional contribution rates is greater than or equal to the cumulative contribution rate constraint are selected as the effective dimensional set. The output dimension pruning rule of the factor encoder is defined as follows: prune other dimensions in the quantitative coding factor other than the effective dimension set.

[0040] In this embodiment of the application, firstly, based on the completed factor encoder, a training verification record is obtained, which includes the dimensional contribution rate distribution of each batch of quantitative coding factors and the corresponding central level term values ​​during the training process.

[0041] Secondly, the training and validation records are analyzed to extract the dimensional contribution rate distribution corresponding to the central level terms, and arranged by the encoded dimensional number to form a contribution rate sequence.

[0042] Specifically, the serialization dimension contribution rate distribution arranges each dimension in descending order of contribution rate and combines it with a preset cumulative contribution rate constraint, usually set in the range of 85% to 95%. The top N dimensions whose sum of dimension contribution rates is greater than or equal to the cumulative contribution rate constraint are selected as the effective dimension set. This constraint ensures that the retained dimensions can explain the main variation information of the original data, while eliminating noisy dimensions with low contribution.

[0043] Furthermore, the output dimension pruning rule of the factor encoder is defined as follows: prune dimensions other than the effective dimension set in the quantitative coding factor. Specifically, this is achieved by setting a dimension mask matrix at the output of the autoencoder layer, setting the output of neurons corresponding to ineffective dimensions to zero, or directly pruning them from the network architecture. The pruned quantitative coding factor dimension is reduced from the original M dimensions to N dimensions, further reducing the input complexity of the subsequent lithology adaptive correction model, while maintaining the integrity of geological information by retaining high-contribution dimensions. The pruned quantitative coding factor is then deeply concatenated with the quantitative lithology factor constructed from lithology indicator elements to form the enhanced quantitative lithology factor that is finally input into the lithology adaptive correction model. This factor integrates the physical interpretability of explicit lithology indicators with the data-driven representation of implicit geological features.

[0044] Further, the corrected signal data and the quantitative lithology factor are input into the lithology adaptive correction model to obtain lithology-corrected equivalent hydrocarbon signal data. The lithology adaptive correction model is configured to dynamically select the correction path based on lithological complexity, including: Based on the quantitative lithology factors, the lithology volatility index is calculated, and the lithology uniformity of the target well section is determined based on the lithology volatility index. If it is determined to be a single lithological region, the first correction path is invoked, and a machine learning regression model deployed on the end side is used to perform lithological correction on the corrected signal data, and the equivalent hydrocarbon signal data is output. If the region is determined to be a mixed lithological area, the second correction path is invoked to perform lithological correction on the corrected signal data and output equivalent hydrocarbon signal data. The second correction path is configured as a hierarchical architecture that combines cloud-based meta-learning pre-training with edge-side incremental fine-tuning.

[0045] In this embodiment, firstly, a lithological volatility index is calculated based on quantitative lithological factors, and the sliding window standard deviation method is used to assess the degree of lithological variation with depth. Specifically, the window length is set to a preset number of sampling points, and the weighted average of the standard deviations of each dimension of the quantitative lithological factors within the window is calculated as the lithological volatility index. The weights are adaptively determined based on the contribution of each dimension to lithological classification. When the lithological volatility index is lower than a preset single-lithology threshold, the target well section is determined to be a single-lithological region; when the index is higher than or equal to the threshold, it is determined to be a mixed multi-lithological region.

[0046] Secondly, for a single lithological region, the first correction path is invoked. This path employs a lightweight machine learning regression model deployed on the end-side, specifically an ensemble model based on gradient boosting decision trees. The model input consists of the concatenated features of the corrected signal data and quantitative lithological factors, and the output is the equivalent hydrocarbon signal data after lithological correction. End-side deployment ensures low-latency response, meeting the requirements of real-time logging. The model obtains the mapping relationship between hydrocarbon signals and lithological factors under single lithological conditions through offline training, and directly calls pre-stored model parameters to complete the correction calculation during online inference.

[0047] For areas with mixed lithologies, a second correction path is invoked. This path employs a hierarchical architecture combining cloud-based meta-learning pre-training and end-side incremental fine-tuning. In the cloud-based meta-learning pre-training stage, a meta-learner is constructed based on full data from multiple historical wells. A model-independent meta-learning algorithm is used to learn initialization parameters that rapidly adapt to different lithological combinations. In the end-side incremental fine-tuning stage, the initial model obtained from the meta-learning pre-training is deployed to the logging site. Online incremental learning is performed on the real-time data stream of the current well section, dynamically adjusting model parameters to adapt to local lithological characteristics. This hierarchical architecture balances the advantages of cloud computing power with the real-time requirements of the end-side, achieving rapid convergence under small sample conditions through meta-learning and adaptive adaptation of the model to the geological specificities of the target well section through incremental fine-tuning.

[0048] Specifically, when a region is identified as having mixed lithology, a second correction path is invoked. This second correction path is configured as a hierarchical architecture combining cloud-based meta-learning pre-training and edge-side incremental fine-tuning, including: The cloud-based system constructs a meta-learning task set based on historical analysis data from multiple wells and trains a deep neural network as a base model using a meta-learning algorithm. The base model is sent from the cloud to the end side, and the sample analysis data of the well section that has been drilled in the current well is obtained to perform incremental fine-tuning on the base model to obtain a field-specific calibration model; The corrected signal data and the quantitative lithology factor are input into the field-specific correction model, and the lithology-corrected equivalent hydrocarbon signal data are output.

[0049] In this embodiment, firstly, a meta-learning task set is constructed in the cloud based on historical analysis data from multiple wells. This task set includes typical well sections from different sedimentary basins, stratigraphic levels, and lithological combinations. Each task is defined as learning the mapping relationship between corrected signal data and equivalent hydrocarbon signal data under a specific lithological combination from a small number of samples. A deep neural network is trained as the base model using a model-independent meta-learning algorithm. This algorithm simulates multi-task scenarios during the meta-training phase, learning a set of initialization parameters sensitive to task changes, enabling the model to quickly adapt to new tasks with only a small number of gradient updates. Specifically, the base model adopts a hybrid architecture of a one-dimensional convolutional neural network with residual connections and an attention mechanism to capture the local features and long-range dependencies of hydrocarbon signal data.

[0050] Secondly, the trained base model is distributed from the cloud to the end-side logging equipment. The end-side equipment acquires sample analysis data of the drilled sections of the current well, including corrected signal data, quantitative lithological factors, and the true values ​​of equivalent hydrocarbon content from laboratory analysis, forming an incremental fine-tuning dataset. An incremental learning strategy based on an experience playback mechanism is used to fine-tune the base model. This strategy maintains a fixed-capacity sample buffer, storing a mixed set of historical representative samples and newly added samples. By constraining the difference between the output of the new model and the output of the old model, it ensures that the model retains learned knowledge while adapting to new lithological characteristics. An adaptive learning rate is set during fine-tuning. In the initial stage, a larger learning rate is used to quickly converge to a local optimum. In the later stages, the learning rate is reduced to finely adjust the parameters. For example, the initial learning rate is set to 0.001, and then gradually decays to 0.0001 until the validation set loss no longer decreases after five consecutive rounds, at which point fine-tuning stops, resulting in a field-specific calibration model optimized for the current well's geological conditions.

[0051] Finally, the real-time acquired corrected signal data and quantitative lithology factors are feature-aligned and stitched together, and then input into a dedicated field correction model. The model outputs lithology-corrected equivalent hydrocarbon signal data through forward propagation calculation. This data eliminates the systematic influence of lithological differences on hydrocarbon detection signals, achieving comparability and quantification of free hydrocarbon content under different lithological conditions. The end-side inference process employs model quantization and operator fusion techniques to optimize the computation graph, compressing 32-bit floating-point weights into 8-bit integers, reducing inference latency to milliseconds while maintaining prediction accuracy.

[0052] S40: Input the lithology-corrected equivalent hydrocarbon signal data into the inversion model to obtain free hydrocarbon content prediction data, and perform cross-validation and output based on the gas logging data of the target well section.

[0053] In this embodiment of the application, an inversion model based on the fusion of physical mechanism and data-driven approach is constructed. This model takes equivalent hydrocarbon signal data as input, establishes the coupling relationship between the thermal evaporation kinetic equation of hydrocarbon substances and the nonlinear mapping of neural network, outputs a quantitative prediction value of free hydrocarbon content, and performs cross-validation based on the gas logging data of the target well section.

[0054] Specifically, step S40 in the method includes: Acquire gas logging data for the target well section, wherein the gas logging data includes at least total hydrocarbon data and component ratio data calculated based on hydrocarbon components; The trend consistency of the predicted free hydrocarbon content data and the total hydrocarbon data is checked, and the correlation coefficient within a preset sliding window is calculated. If the correlation coefficient is lower than a preset threshold, a trend anomaly alarm is triggered. Based on prior logic rules, the fluid property logic consistency check is performed on the free hydrocarbon content prediction data and the component ratio data. If the fluid type indicated by the component ratio contradicts the oil content level corresponding to the free hydrocarbon content prediction data, a fluid type contradiction alarm is triggered. If both the trend consistency check and the fluid property logic consistency check pass, the predicted free hydrocarbon content data will be output.

[0055] In this embodiment of the application, firstly, gas logging data of the target well section is acquired, including total hydrocarbon data and component ratio data calculated based on hydrocarbon components. The total hydrocarbon data reflects the continuous change of total hydrocarbon concentration in the drilling fluid, and the component ratio data includes derived indicators such as relative methane content and light-heavy hydrocarbon ratio, which are used to characterize fluid properties and maturity characteristics.

[0056] Secondly, the trend consistency of the free hydrocarbon content prediction data and the total hydrocarbon data is checked. The sliding window length is set to cover the preset depth range, and the Pearson correlation coefficient of the two sets of data sequences within the window is calculated. This correlation coefficient quantifies the degree of coordinated change between the geochemical logging inversion results and the gas logging response. If the correlation coefficient is lower than the preset threshold, such as 0.6, it indicates that there is a systematic deviation between the free hydrocarbon content prediction data and the gas logging response, which may be due to local overfitting of the lithology correction model or abnormal response of the gas logging system. At this time, a trend anomaly alarm is triggered and the data tracing process is started to investigate the source of the anomaly.

[0057] Secondly, combining prior logic rules, a fluid property logic consistency check is performed on the free hydrocarbon content prediction data and component ratio data. First, a mapping relationship between component ratios and fluid types is established. For example, a high relative methane content usually indicates gas layer or dry gas characteristics, while an increased light-heavy hydrocarbon ratio corresponds to an oil layer or condensate gas layer. If there is a contradiction between the fluid type indicated by the component ratio and the oil-bearing level corresponding to the free hydrocarbon content prediction data, such as a high free hydrocarbon content prediction corresponding to high dry gas characteristics, a fluid type contradiction alarm is triggered, indicating possible drilling fluid mixing interference, formation fluid heterogeneity, or model prediction bias.

[0058] Finally, if both the trend consistency check and the fluid property logic consistency check pass, the reliability of the free hydrocarbon content prediction data is confirmed, and the data is output as a quantitative evaluation result of the free hydrocarbon content in the target well section. The output data includes the free hydrocarbon content curve with continuous depth distribution, the corresponding uncertainty interval estimate, and the quality label.

[0059] In summary, compared with existing technologies, this application has achieved accurate quantitative analysis of the free hydrocarbon content of shale oil by constructing a multi-level data fusion and intelligent correction system.

[0060] This application provides a method for analyzing free hydrocarbon content in shale oil through geochemical logging. First, by simultaneously acquiring raw geochemical logging signal data and elemental logging data, it achieves collaborative acquisition of multi-source geochemical information, overcoming the limitations of traditional methods that rely solely on pyrolysis chromatographic signals. Second, based on a lithology-adaptive correction model constructed using quantitative lithology factors, it dynamically selects the correction path according to lithological complexity. For single-lithology regions, a lightweight end-side model provides rapid response; for mixed-lithology regions, a hierarchical architecture combining cloud-based meta-learning and end-side incremental fine-tuning is employed, effectively solving the technical bottleneck of fixed correction coefficients being unable to adapt to frequent vertical lithological alternations. Third, by mining non-lithological indicator elements through a factor encoder, quantitative coding factors are extracted and integrated into the lithological characterization system, enhancing the model's ability to perceive potential geological features. Finally, gas logging data is introduced for dual cross-validation of trend consistency and fluid property logical consistency, constructing a multi-verification mechanism for the prediction results, improving the reliability and geological applicability of free hydrocarbon content prediction.

[0061] Through the above technical solutions, this application realizes a paradigm shift in shale oil geochemical logging free hydrocarbon content analysis from static empirical correction to dynamic intelligent sensing, improving the accuracy and stability of free hydrocarbon content prediction under complex lithological conditions, and providing technical support for shale oil resource evaluation.

[0062] Example 2, as Figure 2As shown, based on the same inventive concept as the shale oil geochemical logging free hydrocarbon content analysis method provided in Embodiment 1, this application embodiment also provides a smart sensor for shale oil geochemical logging free hydrocarbon content analysis, including: The information acquisition module 11 is used to acquire the original geochemical logging signal data of the target well section and simultaneously acquire elemental logging data, wherein the elemental logging data includes at least the content data of lithological indicator elements. Data correction module 12 is used to preprocess the original geochemical logging signal data to obtain corrected signal data; Model training module 13 is used to determine quantitative lithology factors based on the elemental logging data, and input the corrected signal data and the quantitative lithology factors into the lithology adaptive correction model to obtain lithology-corrected equivalent hydrocarbon signal data. The lithology adaptive correction model is configured to dynamically select the correction path based on the lithological complexity. The result output module 14 is used to input the lithology-corrected equivalent hydrocarbon signal data into the inversion model to obtain free hydrocarbon content prediction data, and to perform cross-validation and output based on the gas logging data of the target well section.

[0063] In one embodiment, the data correction module 12 is specifically used for: Combining a preset sliding window and the local outlier factor method, the original geochemical logging signal data is traversed to identify outliers, and the outliers are replaced with the median value under the sliding window. Adaptive baseline fitting is performed on the replaced original geochemical logging signal data, and the baseline fitting result is subtracted accordingly to obtain the corrected signal data.

[0064] Furthermore, in one embodiment of the application, determining the quantitative lithology factor based on the elemental logging data includes: The content data of lithological indicator elements are extracted from the elemental logging data and normalized to obtain the normalized quantitative value of each lithological indicator element. The normalized quantitative values ​​of each lithological indicator element are combined into multidimensional time-series data and determined as the quantitative lithological factor.

[0065] Furthermore, determining the quantitative lithological factors based on the elemental logging data also includes: The multidimensional time-series data of the quantified values ​​corresponding to the non-lithological indicator elements in the element logging data are defined as quantitative potential factors. A factor encoder is constructed, comprising an autoencoder layer and a decoder layer. The autoencoder layer takes the quantitative latent factor as input and outputs a quantitative encoded factor. The decoder layer takes the quantitative encoded factor as input and outputs a quantitative reduced factor. Construct a joint loss function, and train the factor encoder based on the joint loss function and the sample element logging data. The joint loss function includes a reconstruction loss term based on the mean square error and a central level term based on the dimensional contribution rate. The quantified values ​​of the portion of the elemental logging data that is not the lithological indicator element are input into the autoencoder layer of the factor encoder, and the quantitative coding factor is added to the quantitative lithological factor.

[0066] Further, in one embodiment, a joint loss function is constructed, and the factor encoder is trained based on the joint loss function and combined with sample element logging data, including: Define the mean square error between the quantitative latent factor and the quantitative reduction factor as the reconstruction loss term; Using the multiple coding dimensions of the quantitative coding factors as analysis dimensions, the principal component contribution rate obtained by principal component analysis of any batch of quantitative coding factor sets is defined as the contribution rate of the dimensions. Based on the dimensional contribution rate distribution formed by the multiple dimensional contribution rates, the concentration degree is calculated, and the reciprocal of the concentration degree is defined as the concentration level term; The joint loss function is obtained by weighting the reconstruction loss term and the central level term. Based on historical element logging data, sample element logging data is obtained, and the factor encoder is trained with the goal of minimizing the value of the joint loss function.

[0067] Further, the corrected signal data and the quantitative lithology factor are input into the lithology adaptive correction model to obtain lithology-corrected equivalent hydrocarbon signal data. The lithology adaptive correction model is configured to dynamically select the correction path based on lithological complexity, including: Based on the quantitative lithology factors, the lithology volatility index is calculated, and the lithology uniformity of the target well section is determined based on the lithology volatility index. If it is determined to be a single lithological region, the first correction path is invoked, and a machine learning regression model deployed on the end side is used to perform lithological correction on the corrected signal data, and the equivalent hydrocarbon signal data is output. If the region is determined to be a mixed lithological area, the second correction path is invoked to perform lithological correction on the corrected signal data and output equivalent hydrocarbon signal data. The second correction path is configured as a hierarchical architecture that combines cloud-based meta-learning pre-training with edge-side incremental fine-tuning.

[0068] In one embodiment, the result output module 14 is specifically used for: Acquire gas logging data for the target well section, wherein the gas logging data includes at least total hydrocarbon data and component ratio data calculated based on hydrocarbon components; The trend consistency of the predicted free hydrocarbon content data and the total hydrocarbon data is checked, and the correlation coefficient within a preset sliding window is calculated. If the correlation coefficient is lower than a preset threshold, a trend anomaly alarm is triggered. Based on prior logic rules, the fluid property logic consistency check is performed on the free hydrocarbon content prediction data and the component ratio data. If the fluid type indicated by the component ratio contradicts the oil content level corresponding to the free hydrocarbon content prediction data, a fluid type contradiction alarm is triggered. If both the trend consistency check and the fluid property logic consistency check pass, the predicted free hydrocarbon content data will be output.

[0069] Further, based on historical elemental logging data, sample elemental logging data is obtained, and the factor encoder is trained with the goal of minimizing the value of the joint loss function. This process also includes: Based on the completed training of the factor encoder, obtain the training validation record; Analyze the training and validation records to obtain the dimensional contribution rate distribution corresponding to the set level item; The dimensional contribution rate distribution is serialized, and combined with the preset cumulative contribution rate constraint, the first N dimensions whose sum of dimensional contribution rates is greater than or equal to the cumulative contribution rate constraint are selected as the effective dimensional set. The output dimension pruning rule of the factor encoder is defined as follows: prune other dimensions in the quantitative coding factor other than the effective dimension set.

[0070] Furthermore, when a region is identified as a mixed lithological area, a second correction path is invoked. This second correction path is configured as a hierarchical architecture combining cloud-based meta-learning pre-training and edge-side incremental fine-tuning, including: The cloud-based system constructs a meta-learning task set based on historical analysis data from multiple wells and trains a deep neural network as a base model using a meta-learning algorithm. The base model is sent from the cloud to the end side, and the sample analysis data of the well section that has been drilled in the current well is obtained to perform incremental fine-tuning on the base model to obtain a field-specific calibration model; The corrected signal data and the quantitative lithology factor are input into the field-specific correction model, and the lithology-corrected equivalent hydrocarbon signal data are output.

Claims

1. A method for analyzing free hydrocarbon content in shale oil geochemical logging, characterized in that, include: Obtain raw geochemical logging signal data of the target well section and simultaneously acquire elemental logging data, wherein the elemental logging data includes at least the content data of lithological indicator elements; The raw geochemical logging signal data is preprocessed to obtain corrected signal data; Based on the elemental logging data, a quantitative lithology factor is determined, and the corrected signal data and the quantitative lithology factor are input into a lithology adaptive correction model to obtain lithology-corrected equivalent hydrocarbon signal data. The lithology adaptive correction model is configured to dynamically select the correction path based on the lithological complexity. The lithology-corrected equivalent hydrocarbon signal data is input into the inversion model to obtain free hydrocarbon content prediction data, and cross-validation and output are performed based on the gas logging data of the target well section. The determination of quantitative lithological factors based on the elemental logging data also includes: The multidimensional time-series data of the quantified values ​​corresponding to the non-lithological indicator elements in the element logging data are defined as quantitative potential factors. A factor encoder is constructed, comprising an autoencoder layer and a decoder layer. The autoencoder layer takes the quantitative latent factor as input and outputs a quantitative encoded factor. The decoder layer takes the quantitative encoded factor as input and outputs a quantitative reduced factor. Construct a joint loss function, and train the factor encoder based on the joint loss function and the sample element logging data. The joint loss function includes a reconstruction loss term based on the mean square error and a central level term based on the dimensional contribution rate. The quantified values ​​of the portion of the elemental logging data that is not the lithological indicator element are input into the autoencoder layer of the factor encoder, and the quantitative coding factor is added to the quantitative lithological factor; The process includes constructing a joint loss function and training the factor encoder based on the joint loss function and sample element logging data, comprising: Define the mean square error between the quantitative latent factor and the quantitative reduction factor as the reconstruction loss term; Using the multiple coding dimensions of the quantitative coding factors as analysis dimensions, the principal component contribution rate obtained by principal component analysis of any batch of quantitative coding factor sets is defined as the contribution rate of the dimensions. Based on the dimensional contribution rate distribution formed by the multiple dimensional contribution rates, the concentration degree is calculated, and the reciprocal of the concentration degree is defined as the concentration level term; The joint loss function is obtained by weighting the reconstruction loss term and the central level term. Based on historical element logging data, sample element logging data is obtained, and the factor encoder is trained with the goal of minimizing the value of the joint loss function. Specifically, the corrected signal data and the quantitative lithology factor are input into a lithology adaptive correction model to obtain lithology-corrected equivalent hydrocarbon signal data. The lithology adaptive correction model is configured to dynamically select a correction path based on lithological complexity, including: Based on the quantitative lithology factors, the lithology volatility index is calculated, and the lithology uniformity of the target well section is determined based on the lithology volatility index. If it is determined to be a single lithological region, the first correction path is invoked, and a machine learning regression model deployed on the end side is used to perform lithological correction on the corrected signal data, and the equivalent hydrocarbon signal data is output. If the region is determined to be a mixed lithological area, the second correction path is invoked to perform lithological correction on the corrected signal data and output equivalent hydrocarbon signal data. The second correction path is configured as a hierarchical architecture that combines cloud-based meta-learning pre-training with edge-side incremental fine-tuning.

2. The method for analyzing free hydrocarbon content in shale oil geochemical logging as described in claim 1, characterized in that, The raw geochemical logging signal data is preprocessed to obtain corrected signal data, including: Combining a preset sliding window and the local outlier factor method, the original geochemical logging signal data is traversed to identify outliers, and the outliers are replaced with the median value under the sliding window. Adaptive baseline fitting is performed on the replaced original geochemical logging signal data, and the baseline fitting result is subtracted accordingly to obtain the corrected signal data.

3. The method for analyzing free hydrocarbon content in shale oil geochemical logging as described in claim 1, characterized in that, Based on the elemental logging data, quantitative lithological factors are determined, including: The content data of lithological indicator elements are extracted from the elemental logging data and normalized to obtain the normalized quantitative value of each lithological indicator element. The normalized quantitative values ​​of each lithological indicator element are combined into multidimensional time-series data and determined as the quantitative lithological factor.

4. The method for analyzing free hydrocarbon content in shale oil geochemical logging as described in claim 1, characterized in that, The lithology-corrected equivalent hydrocarbon signal data is input into the inversion model to obtain free hydrocarbon content prediction data, and cross-validation and output are performed based on gas logging data of the target well section, including: Acquire gas logging data for the target well section, wherein the gas logging data includes at least total hydrocarbon data and component ratio data calculated based on hydrocarbon components; The trend consistency of the predicted free hydrocarbon content data and the total hydrocarbon data is checked, and the correlation coefficient within a preset sliding window is calculated. If the correlation coefficient is lower than a preset threshold, a trend anomaly alarm is triggered. Based on prior logic rules, the fluid property logic consistency check is performed on the free hydrocarbon content prediction data and the component ratio data. If the fluid type indicated by the component ratio contradicts the oil content level corresponding to the free hydrocarbon content prediction data, a fluid type contradiction alarm is triggered. If both the trend consistency check and the fluid property logic consistency check pass, the predicted free hydrocarbon content data will be output.

5. The method for analyzing free hydrocarbon content in shale oil geochemical logging as described in claim 1, characterized in that, Based on historical elemental logging data, sample elemental logging data is obtained, and the factor encoder is trained with the objective of minimizing the value of the joint loss function. Following this, the process further includes: Based on the completed training of the factor encoder, obtain the training validation record; Analyze the training and validation records to obtain the dimensional contribution rate distribution corresponding to the set level item; The dimensional contribution rate distribution is serialized, and combined with the preset cumulative contribution rate constraint, the first N dimensions whose sum of dimensional contribution rates is greater than or equal to the cumulative contribution rate constraint are selected as the effective dimensional set. The output dimension pruning rule of the factor encoder is defined as follows: prune other dimensions in the quantitative coding factor other than the effective dimension set.

6. The method for analyzing free hydrocarbon content in shale oil geochemical logging as described in claim 1, characterized in that, When a region is identified as having mixed lithology, a second correction path is invoked. This second correction path is configured with a hierarchical architecture combining cloud-based meta-learning pre-training and edge-side incremental fine-tuning, including: The cloud-based system constructs a meta-learning task set based on historical analysis data from multiple wells and trains a deep neural network as a base model using a meta-learning algorithm. The base model is sent from the cloud to the end side, and the sample analysis data of the well section that has been drilled in the current well is obtained to perform incremental fine-tuning on the base model to obtain a field-specific calibration model; The corrected signal data and the quantitative lithology factor are input into the field-specific correction model, and the lithology-corrected equivalent hydrocarbon signal data are output.

7. A smart sensor for analyzing free hydrocarbon content in shale oil geochemical logging, characterized in that, A method for analyzing free hydrocarbon content in shale oil geochemical logging as described in any one of claims 1-6, comprising: The information acquisition module is used to acquire the original geochemical logging signal data of the target well section and simultaneously acquire elemental logging data, wherein the elemental logging data includes at least the content data of lithological indicator elements. The data correction module is used to preprocess the raw geochemical logging signal data to obtain corrected signal data; The model training module is used to determine the quantitative lithology factor based on the elemental logging data, and input the corrected signal data and the quantitative lithology factor into the lithology adaptive correction model to obtain the lithology-corrected equivalent hydrocarbon signal data. The lithology adaptive correction model is configured to dynamically select the correction path based on the lithological complexity. The result output module is used to input the lithology-corrected equivalent hydrocarbon signal data into the inversion model to obtain free hydrocarbon content prediction data, and to perform cross-validation and output based on the gas logging data of the target well section.

Images