Hydrocarbon geological intelligent sampling while drilling method for mud logging

Abstract

The application relates to the technical field of oil and gas exploration, in particular to a rock debris logging method for intelligent sampling of oil and gas geology while drilling. The method comprises the following steps: collecting rock debris samples in real time according to a preset initial sampling interval, and performing primary cleaning on the collected rock debris samples; acquiring multi-source data from the cleaned wet rock debris; drying the cleaned wet rock debris to obtain dry rock debris; calculating a geological variability index according to rock debris image data; generating a sampling control instruction according to a comparison result of the geological variability index and a preset threshold value; dynamically adjusting the sampling density of subsequent rock debris; and when the geological variability index exceeds a preset variability threshold value, setting a high-density sampling mode with a sampling density greater than the initial sampling interval. The method realizes optimized allocation of sampling resources, fuses image features and engineering parameters through a double-channel hybrid model, reduces the dependence of a new exploration area on labeled data through a transfer learning and collaborative learning mechanism, and realizes continuous evolution of the model.

Description

Technical Field

[0001] This application relates to the field of oil and gas exploration technology, and in particular to a cuttings logging method for intelligent sampling of oil and gas geology during drilling. Background Technology

[0002] Cuttings logging is the most basic, widespread, and essential geological exploration technique in oil and gas exploration and development. During drilling, cuttings generated by the drill bit breaking up formation rocks are returned to the surface with drilling fluid. By collecting and analyzing cuttings samples at a vibrating screen, information such as the lithology and hydrocarbon content of underground formations can be obtained in real time, providing crucial data for geological guidance and reservoir evaluation.

[0003] Traditional cuttings logging relies heavily on manual operation, which presents the following technical challenges: First, manual sand retrieval is time-consuming, resulting in large errors in sampling depth and difficulty in accurately identifying the location of underground strata; second, the description of cuttings is highly subjective, with different geologists potentially providing different descriptions of the same bag of cuttings, lacking standardization; third, manual operation is labor-intensive, with geologists working long hours in high-temperature and noisy environments near vibrating screens, posing safety hazards; and fourth, significant information loss occurs, as the original state information of cuttings samples cannot be permanently preserved after cleaning and drying.

[0004] In recent years, although some automated cuttings collection equipment has emerged, existing methods still have the following shortcomings: the sampling strategy is fixed and cannot dynamically adjust the sampling density according to formation changes, resulting in resource waste in homogeneous formations and easy loss of key information in complex formations; cuttings analysis relies on single image features and does not make full use of multi-source data such as drilling engineering parameters, thus limiting the accuracy of identification; there is a distribution offset between the source domain and the target domain, which leads to a decrease in the model's cross-regional generalization performance. Summary of the Invention

[0005] The purpose of this application is to provide a cuttings logging method for intelligent sampling while drilling in oil and gas geology, which aims to achieve intelligent and precise cuttings logging through dynamic sampling strategies and transfer learning mechanisms.

[0006] In some embodiments of this application, a cuttings logging method for intelligent sampling while drilling in oil and gas geology is provided, including: Rock cuttings samples are collected in real time at a preset initial sampling interval using an automatic sampling device, and the collected rock cuttings samples are then subjected to primary cleaning. The cleaned wet rock cuttings are subjected to multi-source data acquisition, including rock cuttings image data and drilling engineering parameters; The cleaned wet rock fragments are dried to obtain dry rock fragments; Calculate the geological variability index based on the rock debris image data; A sampling control command is generated based on the comparison result between the geological variability index and the preset threshold. The sampling control command is sent to the automatic sampling device to dynamically adjust the sampling density of subsequent rock cuttings; When the geological variability index exceeds a preset variability threshold, a high-density sampling mode with a sampling density greater than the initial sampling interval is set.

[0007] In some embodiments of this application, the calculation of the geological variability index includes: Extract white light images, ultraviolet light images, and real-time drilling parameters from multi-source data; Different lithological grains are segmented and identified from white light images, and the degree of lithological variation is calculated. The intensity and area of ​​fluorescent spots are extracted from ultraviolet images, and fluorescence anomaly values ​​are calculated. Obtain real-time drilling parameters corresponding to the current sampling depth and calculate the engineering anomaly index; A geological variability index is generated based on the lithological variation, fluorescence anomaly value, and engineering anomaly index.

[0008] In some embodiments of this application, the sampling interval for subsequent rock cuttings is dynamically adjusted, including: When the geological variability index is less than the first threshold, maintain the initial sampling interval; When the geological variability index is not less than the first threshold, the high-density sampling protocol is triggered, and the sampling interval of subsequent consecutive sampling points is shortened to the high-density sampling interval. When the geological variability index exceeds the second threshold, an alarm for abnormal strata is simultaneously sent to the control room.

[0009] In some embodiments of this application, the process of drying the wet rock chips to obtain dry rock chips further includes: The cleaned wet rock fragments are sent to the drying unit for drying treatment; After the dried rock cuttings are spread out, they are sent into the imaging chamber to acquire high-definition white light and ultraviolet light images. Based on the high-resolution white light and ultraviolet light images, lithology identification and oil-bearing rating are performed to generate preliminary analysis results.

[0010] In some embodiments of this application, lithological identification is performed, including: The theoretical lag time is calculated in real time based on drilling fluid discharge and annular cross-sectional area parameters. The theoretical lag time is then corrected to obtain the actual depth offset. The acquisition depth of the cuttings image is aligned and matched with the corresponding depth of the drilling engineering parameters based on the actual depth offset.

[0011] In some embodiments of this application, lithological identification and oil-bearing rating include: Construct a dual-channel hybrid model that includes image feature extraction channels and engineering parameter feature extraction channels; Lithological feature vectors are extracted from high-resolution images using image feature extraction channels; The engineering response feature vector is extracted through the engineering parameter feature extraction channel by using a drilling parameter sequence aligned with the depth of the cuttings image; By integrating the lithological feature vector and the engineering response feature vector, the lithological category, oil-bearing level, and corresponding confidence level are output.

[0012] In some embodiments of this application, the dual-channel hybrid model further includes: Obtain historical data from existing oil fields; Based on the historical data, a pre-trained image feature extraction channel is trained through self-supervised learning. For the oilfield to be explored, the fully connected classification layer of the dual-channel hybrid model is fine-tuned; This aims to reduce the difference in characteristic distribution between existing oil fields and oil fields to be explored.

[0013] Some embodiments of this application also include collaborative learning: Obtain the automatic analysis results and store the manual correction records of the automatic analysis results in the experience playback buffer; At preset intervals, samples are extracted from the experience replay buffer and combined with historical data for training, and the dual-channel hybrid model is updated.

[0014] In some embodiments of this application, the fusion of lithological feature vectors and engineering response feature vectors includes: Calculate the attention weight of the lithological feature vector to the engineering response feature vector; The engineering response feature vector is reorganized according to the attention weights to generate engineering enhancement features; The lithological feature vector is concatenated with the engineering enhancement feature and input into the fully connected layer for classification.

[0015] In some embodiments of this application, the sampling density of subsequent rock cuttings is dynamically adjusted, including: When the high-density sampling mode is triggered, the geological variability index value of the current sampling point is obtained; Based on the correspondence between the geological variability index value and the preset threshold range, a pre-stored sampling density level is matched; Set the sampling interval for subsequent consecutive sampling points according to the sampling density level.

[0016] Compared with existing technologies, the cuttings logging method for intelligent sampling while drilling in oil and gas geology, as described in this application, has the following advantages: A dynamic sampling mechanism based on geological variability index was constructed. Through rapid image analysis, lithological variation, fluorescence anomaly value and engineering anomaly index are calculated in real time. The sampling density is automatically adjusted according to the complexity of the strata. The initial sampling interval is maintained in homogeneous strata to save resources, and a high-density sampling mode is triggered in complex strata to capture key information, thereby achieving optimal allocation of sampling resources.

[0017] A dual-channel hybrid model and depth alignment correction mechanism were established to fuse rock cuttings image features with drilling engineering parameters in a multimodal manner. By correcting the lag time, the depth matching of the two types of data was ensured, which solved the depth misalignment problem caused by rock cuttings lag and improved the accuracy of lithology identification and oil-bearing rating.

[0018] A transfer learning strategy based on self-supervised learning and domain adaptation was introduced. The image feature extraction channels were pre-trained using historical data from existing oilfields. By fine-tuning to reduce the differences in feature distribution, the model can quickly adapt to the geological features of new exploration areas and reduce its dependence on a large number of labeled samples in new exploration areas.

[0019] A collaborative learning mechanism was designed to store the manual correction records of geologists in an experience playback buffer, and to periodically extract samples for incremental training, so as to realize the continuous evolution of the model and form a virtuous cycle of human-machine collaboration.

[0020] In the feature fusion stage, an attention weight mechanism is adopted to calculate the attention weight of lithological features to engineering response features and generate engineering enhancement features, enabling the model to automatically learn the correlation of different modal features and improve the fusion effect of multi-source data.

[0021] In high-density sampling mode, the sampling density is precisely adjusted by matching the pre-stored sampling density level with the geological variability index value, thus avoiding information loss or resource waste caused by simple binarization sampling. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a cuttings logging method for intelligent sampling while drilling in oil and gas geology, as described in some embodiments of this application. Detailed Implementation

[0023] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.

[0024] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0025] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0026] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0027] like Figure 1 As shown in the embodiment of this application, a cuttings logging method for intelligent sampling while drilling in oil and gas geology includes: Rock cuttings samples are collected in real time at a preset initial sampling interval using an automatic sampling device, and the collected rock cuttings samples are then subjected to primary cleaning. Multi-source data was acquired from the cleaned wet rock cuttings, including rock cuttings image data and drilling engineering parameters. The cleaned wet rock fragments are dried to obtain dry rock fragments; Calculate the geological variability index based on rock debris image data; Sampling control instructions are generated based on the comparison results between the geological variability index and the preset threshold. The sampling control command is sent to the automatic sampling device to dynamically adjust the sampling density of subsequent rock cuttings; When the geological variability index exceeds the preset variability threshold, a high-density sampling mode is triggered, and the sampling interval corresponding to this high-density sampling mode is smaller than the initial sampling interval.

[0028] The sampling interval refers to a repetitive calculation unit at a specific time or depth during the drilling process. The sampling interval can be synchronized with the drilling depth sampling interval; for example, it can be set as one sampling interval per meter or half a meter of drilling. When the drilling depth reaches the preset sampling interval, the system performs cuttings sampling. The initial sampling interval is a pre-set regular sampling frequency, such as one bag of cuttings collected every two meters in ordinary exploration areas and one bag per meter in key areas.

[0029] Real-time rock cuttings sampling refers to the use of an automatic sampling device installed at the drilling fluid vibrating screen, which is linked to a depth sensor on the drilling platform. The sampling command is automatically triggered when the drilling depth reaches a preset sampling interval. The sampling device uses a robotic arm or screw conveyor to precisely grab the flowing rock cuttings at the rock cuttings discharge outlet of the vibrating screen and immediately sends them to the primary cleaning unit. Primary cleaning involves flushing the rock cuttings with water at a set pressure while simultaneously removing drilling fluid adhering to the surface through a vibrating screen. The screen aperture is selected based on the expected size of the rock cuttings particles in the formation; for example, a 2 mm aperture screen is typically used to ensure that representative rock cuttings particles are retained.

[0030] Multi-source data acquisition of cleaned wet rock cuttings involves conveying the still slightly damp rock cuttings, after initial cleaning, to the imaging unit via a conveyor belt. Images under white light and ultraviolet light are acquired simultaneously within a very short timeframe (milliseconds). The white light image records the natural color, particle morphology, and particle size of the rock cuttings; the ultraviolet image excites any hydrocarbon components present in the rock cuttings, producing fluorescence, and records the distribution, intensity, and color of the fluorescence. The acquired images have moderate resolution but are acquired quickly, primarily for real-time preliminary analysis. The specific resolution is set to 640×480 pixels or 800×600 pixels. This resolution is sufficient to clearly identify the lithological type of the rock cutting particles and the overall distribution of fluorescent patches, but insufficient to observe the microstructure of the particle surface. This resolution is chosen to control the amount of image data while ensuring rapid imaging speed.

[0031] The multi-source data includes cuttings image data and drilling engineering parameters. Specifically, the cuttings image data includes white light images and ultraviolet light images. The drilling engineering parameters include real-time drilling parameters and engineering calculation parameters. The real-time drilling parameters include dynamically acquired data such as drilling time, torque, pump pressure, and total hydrocarbons. The engineering calculation parameters include drilling fluid discharge, annular cross-sectional area, and hysteresis time.

[0032] The drying process involves conveying the wet rock cuttings, after multispectral rapid imaging, to an infrared drying unit via a conveyor belt. Infrared radiation heating is used to rapidly evaporate surface moisture from the rock cuttings. The drying temperature is controlled between 60℃ and 80℃, and the drying time is automatically adjusted according to the amount of rock cuttings (typically 30 to 60 seconds) until the rock cuttings reach a constant weight (mass change rate less than 1%). The dried rock cuttings should retain intact particles, show no overheating discoloration, and exhibit no loss due to the volatilization of oil and gas components.

[0033] Primary cleaning involves using a clean water spray system at a pressure of 0.2–0.5 MPa to wash the cuttings, while simultaneously using a vibrating screen with a 2 mm aperture to remove drilling fluid, mud, and fine impurities adhering to the cuttings surface. During cleaning, the water flow direction is opposite to the cuttings transport direction to ensure thorough washing of the cuttings particles. After cleaning, the liquid content of the cuttings is reduced to below 20%, and there is no obvious drilling fluid residue on the surface.

[0034] The calculation of the geological variability index based on rock cuttings image data involves rapid image processing using a lightweight neural network model deployed on edge computing units. The model first segments the image to distinguish regions of different lithological particles, then identifies features such as particle color and texture, and simultaneously extracts features of fluorescent regions from ultraviolet images. Combined with parameters such as drilling time and torque collected in real-time during the drilling operation, the geological variability index is calculated.

[0035] The lightweight neural network model chosen is the MobileNetV3 network, specifically designed for mobile and edge computing devices. This network has approximately 5 million parameters and an inference time of less than 50 milliseconds per image. The network structure includes an initial convolutional layer, multiple bottleneck layers, and a fully connected classification layer. The bottleneck layers employ depthwise separable convolutions to reduce computation and introduce a squeeze excitation module to enhance feature representation. The model is deployed on an NVIDIA Jetson Xavier NX module in the edge computing unit, utilizing its TensorRT acceleration engine for inference acceleration.

[0036] The geological variability index is a dimensionless numerical value used to quantify the degree of variation in the lithological properties of the current sampling point and the significance of hydrocarbon indications. A higher index indicates a more complex formation or more obvious hydrocarbon indications, necessitating more intensive sampling. Calculating the geological variability index requires pre-setting various weights. The weight generation process is as follows: Collect historical data from drilled wells, including lithological variability, fluorescence anomalies, engineering anomaly index, and the corresponding final sampling density decision. Use a logistic regression model to fit this data and learn the optimal weight combination for the sampling decision. The specific steps are: First, organize the historical data. Each sample contains three features: normalized lithological variability, fluorescence anomalies, and engineering anomaly index, and a label indicating whether intensive sampling is required (1 for intensive sampling, 0 for no intensive sampling). Second, train the logistic regression model using a binary cross-entropy loss function, an Adam optimizer, a learning rate of 0.01, and 100 training epochs. Third, after training, obtain the regression coefficients for the three features, assumed to be β1, β2, and β3. Normalize these three coefficients using the softmax function to obtain the weights w1, w2, and w3, where the formula is... 1=1.2, β2=1.5, β3=0.8, then Approximately 3.32, Approximately 4.48, Approximately 2.23, totaling approximately 10.03, with weights of 0.33, 0.45, and 0.22 respectively. Initially, if historical data is unavailable, empirical weights can be used, such as a weight of 0.4 for lithological variation, 0.4 for fluorescence anomalies, and 0.2 for engineering anomaly indices. These weights can be continuously optimized using the aforementioned methods as data accumulates.

[0037] The training samples for the logistic regression model must meet the following selection criteria: each sampling point must have a corresponding final sampling density decision record, and this decision must have been reviewed and confirmed by a geologist. The label determination criteria are: if the sampling point actually underwent encrypted sampling (i.e., the sampling interval is less than the basic sampling interval), the label is 1; if the initial sampling interval is maintained, the label is 0. Before training, the features need to be standardized so that the mean of the three features is 0 and the standard deviation is 1, to eliminate the influence of dimensions. Five-fold cross-validation is used for model training to prevent overfitting.

[0038] The sampling control command is generated by the edge computing unit comparing the calculated geological variability index with a preset first threshold and a second threshold: if the geological variability index is less than the first threshold, a command to maintain the current sampling interval is generated; if the first threshold is less than or equal to the geological variability index and less than the second threshold, a command to switch to light / medium-level encrypted sampling is generated; if the geological variability index is greater than or equal to the second threshold, a command to switch to heavy-level encrypted sampling is generated, and an alarm command for abnormal strata is generated at the same time.

[0039] The instructions are sent to the programmable logic controller of the automatic sampling device via industrial Ethernet or RS485 bus.

[0040] The automatic sampling device is an electromechanical integrated unit installed at the cuttings discharge outlet of the drilling fluid vibrating screen. It includes: a robotic arm or screw conveyor for precisely grabbing cuttings; a dual-channel buffer mechanism (two independent sampling channels that work alternately to eliminate reset gaps); a cleaning screen and spray system; and a depth sensor interface that links with the drilling rig's depth sensor, automatically triggering sampling when the drilling depth reaches a preset sampling interval. The entire device is PLC-controlled, made of corrosion-resistant stainless steel, and has an IP65 protection rating.

[0041] Dynamic adjustment of subsequent sampling density includes the system dynamically modifying the sampling trigger depth interval of the automatic sampling device according to sampling control instructions: in basic mode, it executes according to the initial sampling interval (e.g., 2 meters / packet); in encrypted mode, it executes according to the high-density sampling protocol (e.g., 1 meter / packet, 0.5 meters / packet, or 0.2 meters / packet); the encryption duration depth is positively correlated with the geological variability index (e.g., 5 sampling points when the index ≥ the first threshold, and 10 sampling points when the index ≥ the second threshold); when the geological variability index of 3 consecutive sampling points is lower than the first threshold, the initial sampling interval is automatically restored.

[0042] The preset variability thresholds include a first threshold and a second threshold, determined by statistically analyzing the quantiles of historical drilling data: The first threshold covers the lowest geological variability index value for 85% of the well sections requiring infill drilling, with an initial default value of 0.3; the second threshold covers the geological variability index value for 95% of the well sections requiring infill drilling, with an initial default value of 0.7. After each well is completed, the two thresholds are dynamically adjusted based on the actual infill drilling effect (effective oil and gas show ratio), with an adjustment step of 0.05.

[0043] High-density sampling mode refers to a state where, when the real-time calculated geological variability index exceeds a first threshold, the system automatically shortens the sampling interval to a level smaller than the initial sampling interval. High-density sampling mode is divided into several levels: Lightly dense: the sampling interval is 1 / 2 of the initial interval (e.g., from 2 meters to 1 meter); Mediumly dense: the sampling interval is 1 / 4 of the initial interval (e.g., 0.5 meters); Heavyly dense: the sampling interval is 1 / 10 of the initial interval (e.g., 0.2 meters).

[0044] Different gear levels correspond to different geological variability index ranges. The system automatically matches and executes based on the current index value until the exit conditions are met.

[0045] In the above embodiments, by introducing a geological variability index, adaptive adjustment of the sampling frequency is achieved. This avoids resource waste caused by over-sampling in homogeneous formations while ensuring sufficiently dense samples are obtained in complex formations or oil and gas-bearing sections, thereby improving the representativeness and accuracy of logging data. By specifying the specific parameters of the resolution image, the specific selection of the lightweight network, the weight generation method, and the statistical determination method of the threshold, the scientific nature and repeatability of the geological variability index calculation and dynamic sampling decision-making are ensured, enabling those skilled in the art to accurately implement this technical solution.

[0046] In some embodiments of this application, the calculation of the geological variability index includes: Extract white light images, ultraviolet light images, and real-time drilling parameters from multi-source data; Different lithological grains are segmented and identified from white light images, and the degree of lithological variation is calculated. The intensity and area of ​​fluorescent spots are extracted from ultraviolet images, and fluorescence anomaly values ​​are calculated. Obtain real-time drilling parameters corresponding to the current sampling depth and calculate the engineering anomaly index; A geological variability index is generated based on lithological variation, fluorescence anomaly value, and engineering anomaly index.

[0047] Identifying different lithological particles refers to using image segmentation algorithms to separate each rock fragment in an image and classify it into a preset lithological category based on visual characteristics such as color, texture, and luster, such as sandstone, mudstone, and limestone.

[0048] The training method and specific implementation of the lithological particle segmentation algorithm are as follows: First, a large number of labeled rock debris images are collected, with annotations including the boundary coordinates and lithological category of each particle, such as sandstone, mudstone, and limestone. The U-Net semantic segmentation network is used for training, with a network structure consisting of an encoder and a decoder. The encoder uses a pre-trained MobileNetV3 as the backbone network to extract image features; the decoder restores image resolution through upsampling and skip connections, outputting the lithological category probability for each pixel. During training, data augmentation methods are used to improve generalization ability, including random rotation from 0 to 90 degrees, random horizontal flipping, random brightness adjustment of ±20%, random contrast adjustment of ±20%, and random Gaussian noise with a standard deviation of 0.01. The cross-entropy loss function is used, the optimizer is Adam, the initial learning rate is 0.001, the batch size is 16, and the training epochs are 100. After training, the model is converted to TensorRT format and deployed on edge computing units. During inference, the model outputs the lithology category of each pixel, and merges adjacent pixels of the same category into the same particle through a connected component analysis algorithm, ultimately obtaining the lithology label and boundary of each particle.

[0049] The rock debris images were labeled using image annotation tools such as LabelMe or COCO Annotator, and were completed by experienced geologists. The annotation rules were as follows: each grain was bounded with a polygon and its corresponding lithology category was selected. After annotation, a consistency check was performed. 10% of the samples were randomly selected and re-annotated by another geologist. The intersection-union ratio (IUR) and category consistency of the two annotations were calculated. If the IUR was below 0.7 or the categories were inconsistent, the annotation was re-added. The labeled dataset was divided into training, validation, and test sets in an 8:1:1 ratio.

[0050] After identifying the various lithological particles, the area or quantity proportion of each type of lithological particle is calculated to obtain the lithological composition of the current sampling point. For example, sandstone accounts for 40% and mudstone accounts for 60%.

[0051] Calculating lithological variability involves comparing the lithological composition of the current sampling point with that of the previous one or several sampling points to quantify the degree of change. The specific calculation formula and normalization method for lithological variability are as follows: First, calculate the area proportion of each type of lithological particle at the current sampling point. Assume there are N types of lithology, and the proportions of each type at the current sampling point are P1, P2, and so on, up to P... N The proportions of various lithologies at the previous sampling point were Q1, Q2 to Q... N The degree of lithological variation is calculated by summing the absolute differences, using the formula: the original value of the lithological variation equals the sum of P from i=1 to N. i -Q i The absolute value of the lithological variation. This original value ranges from 0 to 2, where 0 represents completely identical and 2 represents completely different. Min-max normalization is used to map the original value to the 0-1 range. The formula is: normalized lithological variation equals the original lithological variation value divided by 2. The larger the normalized value, the more drastic the lithological variation.

[0052] Extracting the intensity and area of ​​fluorescent spots from ultraviolet (UV) images involves binarizing the UV image to separate the fluorescent regions from the background. The grayscale threshold is determined as follows: A batch of non-fluorescent background rock debris images are collected, and the 95th quantile of their grayscale distribution is used as the threshold. For example, if 100 non-fluorescent images are analyzed, and the grayscale values ​​are mainly concentrated between 10 and 30, the 95th quantile is 45, then the threshold is set to 45. This threshold is recalibrated for each well to eliminate the influence of differences in UV light source intensity between different batches. Pixels with grayscale values ​​greater than the threshold are marked as fluorescent regions, and pixels with grayscale values ​​less than or equal to the threshold are marked as background. Then, the proportion of the total area of ​​all fluorescent spots to the entire image area is calculated as the fluorescence area index, with the formula: Fluorescence area index = Total number of fluorescent pixels / Total number of pixels in the image, ranging from 0 to 1. Simultaneously, the average grayscale value or brightness value of the fluorescent spots is calculated as the fluorescence intensity index. The grayscale values ​​of all fluorescent pixels are extracted, and their average value is calculated, with the formula: Fluorescence intensity index = Average grayscale value of fluorescent region / 255, ranging from 0 to 1. The higher the fluorescence area and intensity, the more oil and gas components are contained in the rock cuttings, and the higher the fluorescence level.

[0053] Calculating fluorescence anomalies involves combining the fluorescence area index and fluorescence intensity index using a weighted summation method. The formula is: Fluorescence Anomaly = Area Weight x Fluorescence Area Index + Intensity Weight x Fluorescence Intensity Index. The area weight and intensity weight can be set to 0.6 and 0.4 respectively, adjusted based on actual results. The fluorescence anomaly value ranges from 0 to 1; a larger value indicates a more significant oil and gas indication. The determination method for area and intensity weights is as follows: Collect a batch of standard rock cuttings samples with known oil-bearing grades, and calculate the fluorescence area index and fluorescence intensity index for each sample. Use a grid search method, traversing the area weights within a range of 0.1 to 0.9 with a step size of 0.1. The intensity weight is equal to 1 minus the area weight. For each weight set, calculate the fluorescence anomaly value for all samples and calculate the correlation coefficient between the fluorescence anomaly value and the known oil-bearing grade. Select the weight combination that maximizes the correlation coefficient as the final weight. If no standard samples are available, empirical values ​​of 0.6 and 0.4 can be used initially, and the weights can be updated after each well is drilled as data accumulates.

[0054] Real-time drilling parameters refer to parameters such as drilling time, torque, pump pressure, and total hydrocarbons measured in gas, extracted from the drilling engineering monitoring system at the same depth as the current sampling depth. Drilling time refers to the time required to drill a unit of footage, reflecting the drillability of the formation; torque reflects the resistance of the drill bit in breaking the rock and is also related to the rock strength. These parameters can indirectly indicate changes in formation lithology.

[0055] The engineering anomaly index is calculated by comparing the drilling parameters at the current sampling point with the neighboring normal trend values ​​and calculating the degree of deviation. Taking drilling time parameters as an example, drilling time data from N points prior to the current sampling point are first taken. The value of N depends on the rate of formation change. For rapidly changing formations, N is taken as a smaller value, such as 5, to quickly respond to changes; for stable formations, N is taken as a larger value, such as 20, to smooth noise. The system defaults to N being 10, which can be automatically adjusted according to the fluctuation of the drilling time curve during actual drilling; the greater the fluctuation, the smaller N becomes. The mean and standard deviation of these N points are then calculated. The original value of the engineering anomaly index = (current drilling time - mean) / standard deviation. This value represents the multiple by which the current drilling time deviates from the normal level and can be positive or negative. The original value is truncated, limited to between -3 and +3, because a value exceeding 3 times the standard deviation is considered highly significant. The truncated value is then mapped to the interval between 0 and 1 using the formula: Engineering Anomaly Index = (truncation value + 3) / 6. If multiple drilling parameters such as drilling time, torque, and total hydrocarbons are considered simultaneously, the anomaly index of each parameter can be calculated separately, and then the average or weighted average can be taken. For example, the torque anomaly index can be calculated using the same method, and the total hydrocarbon anomaly index can be calculated based on gas measurement values. The final engineering anomaly index is the arithmetic mean of the three.

[0056] Generating a geological variability index based on the lithological variability, fluorescence anomaly value, and engineering anomaly index involves normalizing these three indicators and then weighting them according to preset weights. The normalized lithological variability, fluorescence anomaly value, and engineering anomaly index have all been normalized to the range of 0 to 1 using the aforementioned method. The geological variability index = W1 x normalized lithological variability + W2 x normalized fluorescence anomaly value + W3 x normalized engineering anomaly index, and W1 + W2 + W3 = 1. The weights are determined as described in claim 1, and can be set through historical data fitting or empirical analysis.

[0057] In the above embodiments, by fusing image visual features and engineering parameter features, the geological variability index can comprehensively reflect the lithological changes, hydrocarbon content, and engineering response of strata, providing a scientific basis for dynamic sampling decisions and avoiding misjudgments that may be caused by a single indicator. By detailing the training method of the lithological grain segmentation network, the calculation formula for lithological variability, the extraction steps of fluorescence anomalies, the calculation process of the engineering anomaly index, and the normalization methods for each indicator, the transparency and feasibility of the geological variability index calculation are ensured, enabling those skilled in the art to reproduce the core calculation process of this technical solution.

[0058] In some embodiments of this application, the sampling density of subsequent rock cuttings is dynamically adjusted, including: When the geological variability index is less than the first threshold, maintain the initial sampling interval; When the geological variability index is not less than the first threshold, the high-density sampling protocol is triggered, and the density of subsequent consecutive sampling points is increased to the high-density sampling density. When the geological variability index exceeds the second threshold, an alarm for abnormal strata is simultaneously sent to the control room.

[0059] When the geological variability index is less than the first threshold, the initial sampling interval is maintained. The first threshold is the lower limit for initiating encrypted sampling, determined by the statistical method described in claim 1, for example, set to 0.3. When the geological variability index is below 0.3, the system considers the strata to be stable and no encryption is required, and continues to collect one bag every two meters or one meter at the basic density.

[0060] When the geological variability index is not less than a first threshold, a high-density sampling protocol is triggered. The high-density sampling protocol is a set of preset encrypted sampling rules, including the encrypted sampling density value and the number of continuously encrypted sampling points. For example, when the index is between 0.3 and 0.6, the sampling density is increased to one packet per meter, with at least five sampling points continuously observed. The system automatically adjusts the sampling interval instructions of the automatic sampling device to achieve encryption.

[0061] When the geological variability index exceeds the second threshold, an alarm for an abnormal section is simultaneously sent to the control room. The second threshold is higher than the first threshold, for example, 0.7. At this point, not only does the sampling continue to be intensified, but an alarm also pops up on the monitoring interface in the control room, prompting geologists that there may be important oil and gas shows or complex geological conditions in this well section, requiring manual attention and verification.

[0062] The specific settings for the first and second thresholds are determined by statistically analyzing the quantiles of historical drilling data. The specific operational steps are as follows: First, collect complete logging data from all drilled wells, including lithological variation, fluorescence anomaly values, and engineering anomaly indices for each sampling point, and calculate the geological variability index for each point. Second, experienced geologists mark each well section, indicating which sections belong to complex formations or oil and gas-bearing strata requiring in-depth sampling. Third, plot the cumulative distribution curve of the geological variability index, identify the lowest index value covering 85% of the well sections requiring in-depth sampling as the first threshold, and the index value covering 95% as the second threshold. If no historical data is available, initially set the first threshold to 0.3 and the second threshold to 0.7, continuously recording in-depth sampling decisions and actual discoveries during drilling. Update the thresholds after each well is completed, and use a sliding window method to retain data from the 10 most recent wells for statistical analysis.

[0063] The specific parameters and recovery conditions of the high-density sampling protocol are set as follows: The high-density sampling protocol includes three parts: encryption density level, encryption duration depth, and recovery conditions. The encryption density level, as described in claim 10, is matched to different levels based on the geological variability index value. The encryption duration depth is positively correlated with the geological variability index value. When the geological variability index is between the first and second thresholds, the duration depth is 5 sampling points; when the index exceeds the second threshold, the duration depth is extended to 10 sampling points; if the index remains high, the duration depth increases by 5 points each time it exceeds the second threshold, with a maximum of 20 points. The recovery condition is that the geological variability index of 3 consecutive sampling points is lower than the first threshold, or 4 out of 5 consecutive sampling points are lower than the first threshold and none exceed the second threshold. When the recovery condition is met, the system automatically switches back to the initial sampling interval. If a higher threshold is triggered again during the encryption process, the duration depth is recalculated.

[0064] The determination of the sustained depth of the encryption is based on the following: statistical analysis of the average thickness of abnormal sections in historical drilling data. Assuming the statistical results show that the average thickness of a typical abnormal section between the first and second thresholds is approximately 2 meters, corresponding to 5 sampling points at 0.5-meter intervals; and the average thickness of a significantly abnormal section exceeding the second threshold is approximately 5 meters, corresponding to 10 sampling points at 0.5-meter intervals. For persistently high thicknesses, the sustained depth is increased by 5 points (equivalent to an increase of 2 meters) each time the second threshold is exceeded, to accommodate the potential extension of the abnormal section. If historical data is unavailable, empirical values ​​of 5, 10, and 20 can be used initially, and subsequently adjusted based on actual drilling results.

[0065] In the above embodiments, by setting two thresholds and corresponding response mechanisms, a tiered processing of formation anomalies is achieved: general anomalies only require encrypted sampling, while significant anomalies simultaneously notify personnel for intervention. This ensures the efficiency of automated sampling while retaining the crucial step of manual verification, thus improving the reliability of logging operations. By clearly defining the statistical determination method for the thresholds, the specific parameters of the encryption protocol, and the recovery conditions, the dynamic sampling strategy possesses complete decision-making logic, enabling it to operate stably and reliably during actual drilling operations.

[0066] In some embodiments of this application, the wet rock cuttings are dried to obtain dry rock cuttings, and the process further includes: The cleaned wet rock fragments are sent to the drying unit for drying treatment; After the dried rock cuttings are spread out, they are sent into the imaging chamber to acquire high-definition white light and ultraviolet light images. Based on high-resolution white light and ultraviolet light images, lithology identification and oil-bearing rating are performed, and preliminary analysis results are generated.

[0067] Sending the cleaned rock chips into the drying unit for drying means that the wet rock chips, after primary cleaning, continue to enter the infrared drying device along the conveyor belt, where the residual moisture on the surface of the rock chips is quickly removed by infrared radiation heating, restoring them to a dry state.

[0068] The specific power, temperature, and time control of the infrared drying unit are as follows: The infrared drying unit uses a 2 kW infrared radiation heating tube, with a temperature control range of 50℃ to 100℃. The drying time is automatically adjusted according to the amount of rock debris, which is measured by a weight sensor or optical volume sensor on the conveyor belt. The control logic is as follows: when rock debris is detected entering the drying zone, heating is started and timing begins, with a preset target drying time, such as 45 seconds. Heating stops after the time is up. If the amount of rock debris is large, the conveyor belt dwell time in the drying zone can be extended. The drying time is directly proportional to the amount of rock debris. The thickness of the rock debris layer is measured by an optical volume sensor, and the calculation formula is: drying time equals the base time of 45 seconds multiplied by a thickness coefficient. The thickness coefficient is equal to the actual thickness divided by the standard thickness of 2 mm. For example, when the thickness is 3 mm, the coefficient is 1.5, and the drying time is 68 seconds. Temperature control uses a PID algorithm, adjusting the heating power based on feedback from the infrared temperature sensor to maintain a stable temperature at the set value, such as 70℃. Temperature feedback is also provided, and heating is stopped early when the surface temperature of the rock debris reaches 80℃. The drying temperature is usually controlled between 60℃ and 80℃ to avoid excessively high temperatures causing the volatilization of oil and gas components in the rock cuttings or changes in the rock minerals.

[0069] The dried rock chips are evenly spread out and enter the imaging chamber to collect high-definition white light images and ultraviolet fluorescence images. This means that the dried rock chips are evenly distributed on a transparent glass plate by a vibrating feeder or conveyor belt to form a single non-overlapping particle layer, and then sent into the high-resolution imaging chamber.

[0070] The specific camera resolution and lens parameters of the imaging chamber are as follows: The high-definition industrial camera uses a 20-megapixel CMOS sensor with a resolution of 5472×3648 pixels, equipped with a telecentric lens to eliminate perspective distortion and ensure that the particle size in the image is linearly related to the actual size. The lens magnification is 0.5x, the working distance is 200 mm, and the depth of field is 10 mm, ensuring that even single-layered rock debris particles can be clearly imaged. The imaging chamber is equipped with a ring-shaped LED white light source with a color temperature of 5500K, a color rendering index greater than 90, and adjustable light intensity. The ultraviolet light source uses a 365 nm LED array with adjustable power, and the white light source is automatically turned off during illumination. First, high-definition photos are taken under white light to record the true color, particle shape, surface structure, roundness, sorting, and other microscopic characteristics of the rock debris; then, the system automatically switches to the ultraviolet light source to take high-definition ultraviolet fluorescence photos, recording the specific occurrence of fluorescence on each particle, such as intergranular pore fluorescence, crack fluorescence, or particle surface fluorescence, as well as the color and intensity distribution of the fluorescence.

[0071] The illumination uniformity requirement for the imaging chamber is that the difference in illumination intensity at any location within the imaging area should not exceed ±5%. Illumination uniformity calibration is required after each power-on. A standard white board is placed at the imaging position, images are captured, and the grayscale values ​​of each area are analyzed. If the grayscale value difference exceeds 5%, the brightness of the LEDs in each area is automatically adjusted until the requirement is met. The ultraviolet light source also requires uniformity calibration to ensure the accuracy of fluorescence intensity measurements.

[0072] Based on the high-definition white light images and ultraviolet fluorescence images, lithology identification and oil-bearing rating are performed to generate preliminary analysis results. This means that the high-definition images are input into a high-precision deep learning model deployed on an edge computing unit or in the cloud. After training, the model can identify different lithologies and automatically assess the oil-bearing level based on fluorescence characteristics.

[0073] The specific criteria and quantitative basis for classifying oil content levels are as follows: Oil content levels are determined based on the area ratio of fluorescent regions and the fluorescence intensity in the fluorescence image. Fluorescence intensity thresholds are obtained by calibrating standard samples with known oil content levels. Standard rock cutting samples of different levels are collected. The fluorescence intensity of oil-saturated samples is typically between 200 and 255, oil spot samples between 150 and 200, oil trace samples between 100 and 150, fluorescent samples between 50 and 100, and samples with no visible fluorescence are less than 50. The lower limit of each interval is taken as the classification threshold. The grading criteria are as follows: Oil saturation corresponds to a fluorescent area greater than 50% and a fluorescence intensity greater than 200; oil spots correspond to a fluorescent area of ​​10% to 50% and a fluorescence intensity of 150 to 200, or a fluorescent area greater than 50% but a fluorescence intensity of 150 to 200; oil traces correspond to a fluorescent area of ​​1% to 10% and a fluorescence intensity of 100 to 150, or a fluorescent area of ​​10% to 50% but a fluorescence intensity of 100 to 150; fluorescence corresponds to a fluorescent area less than 1% but visible fluorescence, or a fluorescent area of ​​1% to 10% but a fluorescence intensity less than 100; no display corresponds to a region without fluorescence. The model outputs the lithological designation, oil-bearing level, and confidence level for each sampling point. These results constitute the preliminary analysis and are provided for verification by geologists.

[0074] In the above embodiments, by combining primary rapid imaging with subsequent high-definition imaging, both the speed requirements of real-time dynamic sampling and the precision of the final rock cuttings description are met, achieving a balance between efficiency and accuracy. By clearly defining the physical parameters of the drying unit, the optical parameters of the imaging chamber, and the quantitative standards for oil content, the standardization and repeatability of the rock cuttings processing and analysis process are ensured.

[0075] In some embodiments of this application, lithological identification is performed, including: The theoretical lag time is calculated in real time based on drilling fluid discharge and annular cross-sectional area parameters. The theoretical lag time is then corrected to obtain the actual depth offset. The acquisition depth of the cuttings image is aligned and matched with the corresponding depth of the drilling engineering parameters based on the actual depth offset.

[0076] The theoretical lag time, calculated in real time based on drilling fluid displacement and annular cross-sectional area parameters, refers to the time required for cuttings to be carried from the bottom of the well to the surface by the drilling fluid, according to the working principle of the drilling fluid circulation system. This time is called the lag time.

[0077] The specific calculation formulas and units for theoretical lag time are uniformly explained as follows: Theoretical lag time is in minutes, and the formula is: Theoretical lag time equals annular volume divided by drilling fluid displacement. Annular volume is in cubic meters, and the formula is: Annular volume equals annular cross-sectional area multiplied by well depth. Annular cross-sectional area is in square meters, and the formula is: Annular cross-sectional area equals π multiplied by the square of the wellbore radius minus the square of the drill string radius. The wellbore radius and drill string radius need to be obtained in real time based on the actual drill string assembly. The wellbore diameter is obtained from the drilling engineering design documents, but the actual wellbore may have enlargement, which can be corrected in real time using wellbore diameter data from logging-while-drilling. The drill string outer diameter is obtained in real time from the drill string assembly table and is automatically updated each time a single string is connected or a drill string is changed. The system maintains a drill string assembly database and automatically matches the corresponding drill string outer diameter based on the current well depth. Drilling fluid displacement is in cubic meters per minute and is read in real time from the drilling engineering monitoring system. For example, if a well is 1000 meters deep, has a borehole diameter of 215.9 mm and a radius of 0.10795 m, and a drill string outer diameter of 127 mm and a radius of 0.0635 m, then the annular cross-sectional area is equal to 3.14 x (0.10795 m). 2 -0.0635 2 )≈0.024m 2 The annular volume is equal to 0.024 x 1000 = 24 m³. 3 If the drilling fluid discharge rate is 1.2m³ 3 If the drilling fluid displacement is / min, then the theoretical lag time is equal to 24 / 1.2 = 20min. The system acquires drilling fluid displacement data in real time and calculates the annulus volume based on the current well depth, thereby updating the theoretical lag time in real time.

[0078] Correcting the theoretical lag time to obtain the actual depth offset is necessary because factors such as drilling fluid flow and cuttings slippage can cause discrepancies between the theoretical and actual lag times. To address this, a conductivity tracer can be added to the drilling fluid, or a natural gamma detector for cuttings can be installed on the drill string. The actual lag time can be obtained by detecting the time it takes for the tracer to reach the surface or by directly measuring the natural radioactivity of the cuttings. Using the actual lag time to correct the theoretical value yields a more accurate depth offset, indicating the true depth from which the currently collected cuttings originate.

[0079] The specific usage and detection method of the conductivity tracer are as follows: When it is necessary to measure the actual lag time, the system automatically controls the injection device to inject a high-concentration potassium chloride solution, for example, 20% concentration, in a single injection of 50 liters into the drilling fluid. The injection frequency is once every 500 meters or once every 24 hours, whichever comes first. During injection, the solenoid valve is opened, and the metering pump injects 50 liters of solution, while simultaneously recording the injection time T0. A conductivity sensor is installed at the outlet of the vibrating screen to continuously monitor the conductivity of the returned drilling fluid. When the conductivity rises by more than 20% above the baseline, the peak value is recorded. After the peak value, the lag time is automatically calculated. The actual lag time is equal to the time when the conductivity peak occurs, T1, minus the injection time T0. This method can be performed regularly once a day, or once every 500 meters increase in well depth, to correct for the cumulative error of the theoretical lag time.

[0080] The installation method and applicable scenarios of the natural gamma ray detector are as follows: The natural gamma ray detector can be installed in the drill collar near the drill bit to measure the natural gamma radioactivity of the formation. When the formation is broken by the drill bit, the detector records the gamma value at that depth point, and simultaneously records the time. After the cuttings reach the surface, their natural gamma radioactivity has been recorded by the detector. By comparing the downhole measurement time and the surface sampling time, the actual lag time of the cuttings bag can be accurately calculated. This method has high accuracy, but the detector is expensive and requires the support of a measurement-while-drilling data transmission system, making it suitable for key wells with high-precision exploration requirements.

[0081] Aligning the acquisition depth of the cuttings image with the corresponding depth of the drilling parameters based on the actual depth offset involves subtracting the depth difference corresponding to the actual lag time from the sampling depth of the cuttings image to obtain the true formation depth of the cuttings. The depth difference equals the actual lag time multiplied by the current drilling rate. Simultaneously, this true depth is matched with the depth records of drilling parameters such as drilling time and gas logging to ensure that all data used for subsequent analysis corresponds to the same depth point, eliminating depth misalignment caused by cuttings run-through lag. The aligned data is stored in a real-time database with depth as the primary key. Each record includes depth, the corresponding cuttings image file name, lithology identification result, oil-bearing grade, drilling time, torque, total hydrocarbons, pump pressure, etc. All subsequent analyses are based on this aligned dataset to ensure depth consistency.

[0082] The current drilling rate is measured in meters per minute (m / min) and is calculated as follows: Current drilling rate equals the footage advanced at the current sampling interval divided by the corresponding pure drilling time. For example, if the sampling interval is 0.5 meters and the pure drilling time is 10 minutes, then the drilling rate is 0.05 m / min. The actual depth offset equals the actual lag time (in minutes) multiplied by the current drilling rate (in meters per minute), also in meters. For example, if the actual lag time is 20 minutes and the current drilling rate is 0.05 m / min, then the depth offset is 1 meter. The true formation depth of the cuttings is equal to the sampling depth minus the depth offset.

[0083] In the above embodiments, by accurately calculating and correcting the lag time, depth alignment between cuttings data and engineering data was achieved, solving the persistent problem of inaccurate depth positioning in traditional logging and laying a solid foundation for multimodal data fusion analysis. By clarifying the calculation formula for theoretical lag time, the operation method of conductivity tracer, and the application scenarios of natural gamma detectors, depth alignment correction has multiple implementation methods, and those skilled in the art can choose the appropriate method according to actual conditions.

[0084] In some embodiments of this application, lithological identification and oil-bearing rating include: Construct a dual-channel hybrid model that includes image feature extraction channels and engineering parameter feature extraction channels; Lithological feature vectors are extracted from high-resolution images using image feature extraction channels; The engineering response feature vector is extracted through the engineering parameter feature extraction channel by using a drilling parameter sequence aligned with the depth of the cuttings image; By integrating lithological feature vectors and engineering response feature vectors, the lithological category, oil-bearing level, and corresponding confidence level are output.

[0085] Constructing a dual-channel hybrid model that includes image feature extraction and engineering parameter feature extraction channels involves designing a neural network architecture with two parallel branches: one branch specifically processes high-resolution image data, and the other branch processes drilling parameter time-series data aligned with the image depth. The two branches extract high-level features from their respective modalities, which are then fused at an intermediate layer, and finally the classification result is output through a fully connected layer.

[0086] The specific network structure and input / output dimensions of the image feature extraction channel are as follows: ResNet50 is used as the image feature extraction network, with high-resolution white light images and ultraviolet fluorescence images as inputs. After high-resolution image acquisition, the image is first cropped to remove the background area and retain the rock debris distribution area. Then, the image is scaled to 224×224 pixels using bilinear interpolation. The white light image and ultraviolet fluorescence image are normalized to the 0-1 interval respectively, and then concatenated along the channel dimension to form a 224×224×6 input tensor. If only a white light image is available and no ultraviolet image is available (e.g., no fluorescence sample), the ultraviolet channel is filled with 0. The ResNet50 network includes an initial convolutional layer, 16 residual blocks, and a global average pooling layer, ultimately outputting a 2048-dimensional feature vector. To accommodate edge computing, a lightweight version such as ResNet18 can be used, outputting a 512-dimensional feature vector. The network weights are obtained using the transfer learning strategy described in claim 7.

[0087] Extracting lithological feature vectors from high-resolution images through image feature extraction channels refers to inputting high-resolution white light images and ultraviolet fluorescence images into a convolutional neural network. The network automatically learns and outputs a fixed-length feature vector through multiple layers of convolution and pooling operations. This vector encodes the lithological-related features of the image, such as grain texture, color distribution, and fluorescence morphology.

[0088] The specific network structure and input sequence length of the engineering parameter feature extraction channel are as follows: A two-layer LSTM network is used as the engineering parameter feature extraction channel. The input is a time series sequence of drilling parameters aligned with the current cuttings depth. The sequence length is set to 10, meaning it takes drilling parameters from the current depth point and the previous 9 depth points. The parameters at each time point include four parameters: drilling time, torque, total hydrocarbons, and pump pressure, forming a 10×4 input matrix. Sequence sampling is centered on the current depth point, taking 9 points forward and 0 points backward (i.e., only historical data is taken, as future data is unknown). If there are fewer than 9 points before the current depth point, the value of the first point is used to fill the gap. The drilling parameters at each point need to be standardized by subtracting the mean and dividing by the standard deviation. The mean and standard deviation are obtained statistically from historical data. The LSTM hidden layer dimension is set to 128, and the hidden state of the last time step is output as the engineering response feature vector, with a dimension of 128.

[0089] Extracting engineering response feature vectors through the engineering parameter feature extraction channel refers to inputting a segment of drilling parameters near the same real depth point into a time series network. The network captures the patterns of changes in parameters such as drilling time, torque, and total hydrocarbons over time and outputs engineering response feature vectors, which are used to characterize engineering response features related to lithological changes.

[0090] The fusion of lithological feature vectors and engineering response feature vectors refers to combining two feature vectors in a certain way to obtain a fused feature vector that contains both visual and engineering information.

[0091] The specific implementation and dimensionality processing method of feature fusion are as follows: Two feature vectors are fused using a concatenation method, but the engineering response feature vector needs to be mapped to the same dimension as the image feature vector first. Specifically, a fully connected layer is added after the engineering feature extraction channel, mapping the 128-dimensional engineering feature to 2048 dimensions, matching the image feature dimension. Then, the 2048-dimensional image feature and the 2048-dimensional mapped engineering feature are concatenated to obtain a 4096-dimensional fused feature. The fused feature is input into a classification network containing two hidden layers: the first hidden layer has 1024 dimensions, and the second hidden layer has 512 dimensions. The final output nodes represent the number of lithology categories and the number of oil-bearing grades.

[0092] Outputting lithology category, oil-bearing grade, and corresponding confidence level means inputting the fused feature vector into the fully connected layer and the softmax function to obtain the probability distribution of each lithology category and oil-bearing grade. The one with the highest probability is the identification result, and the corresponding probability value is the confidence level.

[0093] The specific meaning and threshold setting method of confidence level are as follows: Confidence level is the maximum probability value of the softmax output layer, ranging from 0 to 1, representing the model's confidence in the current prediction result. The confidence level threshold is determined through performance statistics of the validation set. A batch of labeled validation samples are collected, and the accuracy and recall of the model under different confidence level thresholds are statistically analyzed. The threshold that maximizes the F1 score is selected as the default value. Typically, the confidence level threshold is set to 0.8. When the confidence level of the output result is greater than or equal to 0.8, the result is automatically entered into the database as the final result; when the confidence level is less than 0.8, the system marks the sample as "requiring manual review," highlights it on the interface, and prioritizes reminding geologists to make a manual judgment. After the geologist's review, the record enters the experience playback buffer described in claim 8 for subsequent model optimization.

[0094] The standardized mean and standard deviation are calculated using a sliding window method. All historical data within a 100-meter range prior to the current depth point are used to calculate the mean and standard deviation for each parameter. If the range prior to the current depth point is less than 100 meters, all available historical data are used. This method can adapt to parameter variations in different well sections; for example, the drilling mean of the upper formation may differ from that of the lower formation, and the sliding window method can track these changes in real time. The standardization formula is: standardized value = original value - sliding mean divided by sliding standard deviation.

[0095] It is understood that in the above embodiments, by fusing multimodal data through a dual-channel hybrid model, lithology identification not only relies on the appearance of rock cuttings but also incorporates the response characteristics of the drilling process, thereby improving the accuracy and robustness of identification. This is particularly suitable for situations where rock cuttings have similar appearances but different engineering responses. By clearly defining the specific structure, input and output dimensions, feature fusion method, and confidence threshold settings of the dual-channel hybrid network, those skilled in the art can fully construct and deploy this hybrid model.

[0096] In some embodiments of this application, the dual-channel mixing model further includes: Obtain historical data from existing oil fields; Image feature extraction channels are pre-trained based on historical data through self-supervised learning. For oilfields to be explored, the fully connected classification layer of the dual-channel hybrid model is fine-tuned; To reduce the difference in characteristic distribution between existing oil fields and exploratory oil fields.

[0097] Acquiring historical data from existing oilfields refers to collecting a large number of labeled cuttings images and corresponding drilling data from maturely explored oilfields. These data have been manually verified by geological experts and contain accurate lithology labels and oil-bearing grade labels, forming a high-quality source domain dataset.

[0098] Pre-training image feature extraction channels based on the historical data through self-supervised learning refers to training the image feature extraction network on the source domain data using a self-supervised learning task, enabling it to learn general rock image representations without relying on labels.

[0099] The specific tasks and implementation methods of self-supervised learning are as follows: Self-supervised pre-training is performed using the SimCLR framework from contrastive learning. The specific steps are as follows: First, a batch of images is randomly selected from existing historical rock cuttings images of oilfields. Each image undergoes two different random data augmentations to obtain two augmented views. The data augmentation parameters are: random cropping area is 0.08 to 1.0 times the original image size; random color jittering with maximum brightness adjustment of 0.4, maximum contrast of 0.4, maximum saturation of 0.2, and maximum hue of 0.1; random Gaussian blur with a standard deviation of 0.1 to 2.0; and random grayscale probability of 0.2. Second, the two augmented views are input into the image feature extraction network to obtain two feature vectors. Third, the cosine similarity of the two feature vectors is calculated using the contrastive loss function InfoNCE, which aims to make the features of the two views of the same image as close as possible and the features of different images as far apart as possible. The formula for the InfoNCE loss function is... Where sim represents the cosine similarity, and τ is the temperature coefficient, typically set to 0.1. The fourth step is backpropagation to update the network parameters. After pre-training, the image feature extraction network learns to distinguish the general features of different rock debris images without any annotation information. Pre-training is usually performed on a cloud server, utilizing GPU clusters for acceleration.

[0100] The temperature coefficient τ controls the degree to which the contrastive loss function focuses on difficult samples. A smaller τ indicates that the model pays more attention to negative samples with high similarity to positive samples; a larger τ indicates that the model focuses more evenly on all negative samples. Experience shows that a τ value between 0.1 and 0.5 performs well in image contrastive learning. This approach uses a value of 0.1, and preliminary experiments on the standard image dataset ImageNet have verified that this value enables the model to achieve good feature representation quality on rock cuttings images. If the performance is unsatisfactory in subsequent applications in new oil fields, grid search optimization can be performed within the range of 0.05 to 0.5.

[0101] For the oilfield to be explored, fine-tuning the fully connected classification layer of the dual-channel hybrid model means transferring the pre-trained feature extraction channels to the new oilfield, training the fully connected classification layer only on a small amount of labeled data from the new oilfield, while keeping the parameters of the feature extraction layer basically unchanged or making only minor adjustments. Since the feature extraction layer has already learned common rock features, a small amount of new data allows the model to quickly adapt to the geological characteristics of the new oilfield.

[0102] During the fine-tuning process, domain adaptation technology is adopted to enable the feature extractor to learn features that retain source domain knowledge and align with the distribution of the target domain. This reduces the distribution shift caused by differences in appearance such as rock color and texture between the two oil fields and improves the model's generalization ability in new oil fields.

[0103] The specific network structure and training method of the domain adaptation technique are as follows: Domain adaptation is achieved by training a Domain Adversarial Neural Network (DANN). The network structure consists of three parts: a feature extractor (i.e., a pre-trained image feature extraction network), a label classifier for lithology classification, and a domain discriminator to distinguish whether features come from the source domain or the target domain. During training, the source domain data already has oilfield lithology labels, while the target domain data is from a newly explored area and has no labels or only a few labels. There are three training objectives: first, to minimize the classification loss of the label classifier in the source domain; second, to maximize the classification loss of the domain discriminator, even if the feature extractor deceives the domain discriminator; and third, to minimize the classification loss of the domain discriminator, even if it can distinguish between the source and target domains. Through this adversarial training, the feature extractor learns features that maintain classification ability and are robust to domain changes. The training adopts an alternating update strategy: in each training batch, the feature extractor is fixed first, and the domain discriminator is updated one step; then the domain discriminator is fixed, and both the feature extractor and the label classifier are updated one step. The total number of training rounds is 50, with an initial learning rate of 0.001, which decays to 0.1 times the original rate every 20 rounds.

[0104] In the above embodiments, the reliance on large amounts of labeled data in new exploration areas is significantly reduced through transfer learning strategies, enabling the intelligent logging system to be quickly put into use in the early stages of drilling and continuously optimized as data accumulates, demonstrating strong practicality and economy. By elaborating on the specific tasks of self-supervised learning and the network structure and training methods of domain adaptation technology, those skilled in the art can reproduce the transfer learning process and achieve rapid model adaptation in new exploration areas.

[0105] Some embodiments of this application include collaborative learning: Obtain the results of automatic analysis and store the records of manual corrections to the results of automatic analysis in the experience playback buffer; At preset intervals, samples are drawn from the experience replay buffer and combined with historical data for training.

[0106] The process of acquiring automated analysis results and storing manual corrections therein in an experience playback buffer involves displaying each automatically generated lithology identification and oil-bearing grade result on the geologist's workstation for review. If the geologist deems the AI's judgment correct, they can click "confirm"; if they believe it is incorrect, they can manually correct it, for example, changing "mudstone" to "sandy mudstone" or "fluorescence" to "oil stain." Each correction record, along with the corresponding image data, engineering data, and the original AI output, is stored in a fixed-size buffer—the experience playback buffer.

[0107] The specific size and management strategy of the experience replay buffer are as follows: The experience replay buffer adopts a circular queue structure with a fixed size of 2000 records. Each record includes high-resolution image data, the corresponding drilling parameter sequence, AI-automated analysis results of lithology category and oil-bearing level, and geologist's manual correction results. When the buffer is full, new records automatically overwrite the oldest records. Records in the buffer are stored in depth order to facilitate sampling by depth interval. To balance the samples of different types, lithology categories need to be resampled, using a combination of oversampling and undersampling. The number of samples for each lithology category in the buffer is counted, and the target number is set to 1.5 times the average number of each category. For categories with fewer samples than the target number, existing samples are randomly copied for oversampling; for categories with more samples than the target number, some samples are randomly discarded for undersampling. This ensures that the samples of each category are balanced during training.

[0108] At preset intervals, samples are drawn from the experience replay buffer and combined with historical data for training. This means that at certain depth intervals, such as every 100 meters or time intervals, such as every 24 hours, the system randomly draws a batch of manually corrected samples from the buffer and mixes them with the previous historical training data to incrementally train the dual-channel hybrid model.

[0109] The specific hyperparameter settings for incremental training are as follows: Incremental training is triggered once every 24 hours, or once every 100 meters of drilling. During training, 256 samples are randomly selected from the experience replay buffer and mixed with the original training data at a 1:1 ratio to form a training batch. A small learning rate, such as 0.0001, is used, approximately one-tenth of the initial training learning rate. The optimizer remains Adam, with a batch size of 32 and 3 training epochs to avoid overfitting to new samples. During training, the first few layers of the feature extractor are frozen, and only the last few layers and the classification layer are updated to maintain the learned general features. After training, the accuracy of the model on the most recent 100 new samples is verified. If the accuracy improves, a new model is deployed; otherwise, the original model is maintained. The criteria for judging the accuracy improvement are: the accuracy of the new model on the most recent 100 new samples is at least 1 percentage point higher than that of the old model, and the decrease in the number of categories does not exceed 2. At the same time, it is necessary to verify that the accuracy of the new model on the historical test set does not decrease by more than 2 percentage points to avoid catastrophic forgetting.

[0110] In the above embodiments, by introducing experience replay and incremental training, continuous learning through human-machine collaboration is achieved: each correction by the geologist becomes nourishment for model evolution, the model becomes more accurate with use, gradually approaching expert level, while reducing repetitive work for the geologist, forming a virtuous cycle. By clearly defining the size and management strategy of the experience replay buffer, as well as the hyperparameter settings for incremental training, the collaborative learning mechanism can operate stably, achieving continuous model optimization.

[0111] In some embodiments of this application, the fusion of lithological feature vectors and engineering response feature vectors includes: Calculate the attention weights of the lithological eigenvectors on the engineering response eigenvectors; The engineering response feature vector is reorganized based on attention weights to generate engineering enhancement features; The lithological feature vector is concatenated with the engineering enhancement feature and then input into the fully connected layer for classification.

[0112] Calculating the attention weight of the lithological feature vector to the engineering response feature vector means using an attention mechanism to dynamically evaluate the importance of each dimension in the engineering response features for the current lithological identification task.

[0113] The specific calculation steps and formulas for attention weights are as follows: A scaled dot product attention mechanism is used. Let the lithological feature vector be Q, and the dimension d... k The engineering response feature vectors are K and V, both with dimension d. kQ is the lithological feature vector output from the image feature extraction channel, and K and V are the engineering response feature vectors output from the engineering parameter feature extraction channel, after undergoing the same linear transformation. The linear transformation uses a fully connected layer to map the original engineering feature vectors to the same dimension as Q. The transformation matrix and the attention mechanism are trained end-to-end. The attention score is calculated using the following formula: The dimensions of Q and K must be consistent. If the dimensions of the engineering features differ from those of the image features, they must first be mapped to the same dimension using a fully connected layer. After obtaining the score vector, it is normalized to weights using the softmax function, with the formula Wi = N is the dimension of the engineering feature vector. The final engineering enhancement feature is equal to the sum of weights i multiplied by V from i=1 to N. i This means that the engineering feature vectors are summed by weights according to attention weights.

[0114] To ensure consistency in the dimensions of Q, K, and V, a fully connected layer is added after the engineering response feature vector is output from the engineering parameter feature extraction channel. This fully connected layer maps the vector to the same dimensions as the image feature vector. Assuming the image feature vector is 2048-dimensional and the engineering response feature vector is 128-dimensional, the fully connected layer's input is 128-dimensional, and its output is 2048-dimensional. The weight matrix of this fully connected layer is trained end-to-end along with the attention mechanism. After training, both K and V are this mapped 2048-dimensional vectors, consistent with the dimension of Q. The dimensions d of Q, K, and V are... k That is, 2048.

[0115] To illustrate with a specific numerical example: Suppose the lithological feature vector Q is 2048-dimensional, and the engineering feature vectors K and V are also mapped to 2048 dimensions. Calculating the dot product of the transpose of Q and K yields 2048 scores. Assume the score in the 100th dimension is the largest, 5.0, and the scores in other dimensions are between -2 and 2. Divide by... After 45.3, the range of possible scores narrows, for example, 5.0 / 45.3. 0.11. After softmax, the weight of the 100th dimension is approximately 0.15, while the weights of the other dimensions are distributed around 0.04. The total weight of all 2048 dimensions is 1, with the 100th dimension at 0.15 and the other 2047 dimensions averaging approximately 0.000415 each, for a total of 0.15 + (2047). 0.000415) 1.0. This aligns with the properties of attention weights, where a few key dimensions receive higher weights, while most dimensions have weights close to 0. Multiplying the weights by the corresponding dimensions of V and summing the results yields the engineering enhancement feature, where lithological information is amplified and irrelevant information is suppressed.

[0116] Reorganizing the engineering response feature vector according to the attention weights to generate engineering enhanced features means weighting and summing the attention weights with the engineering response feature vector to obtain a new feature vector that highlights engineering information related to the current lithological characteristics and suppresses irrelevant noise.

[0117] The process of concatenating the lithological feature vector with the engineering enhancement feature and inputting it into the fully connected layer for classification involves concatenating the original lithological feature vector and the attention-weighted engineering enhancement feature along the feature dimension to form a higher-dimensional fusion feature, which is then fed into the fully connected layer for the final classification decision.

[0118] In the above embodiments, an attention mechanism is used to dynamically fuse image features and engineering features, allowing the model to adaptively select the most relevant engineering information to assist in lithology determination. This is more flexible and effective than simple splicing or addition, thus improving the fusion effect. By detailing the calculation steps and examples of attention weights, those skilled in the art can accurately implement the attention mechanism for feature fusion.

[0119] In some embodiments of this application, the sampling density of subsequent rock cuttings is dynamically adjusted, including: When the high-density sampling mode is triggered, the geological variability index value of the current sampling point is obtained; Based on the correspondence between the geological variability index value and the preset threshold range, the pre-stored sampling density level is matched; Set the sampling interval for subsequent consecutive sampling points according to the sampling density level.

[0120] When the high-density sampling mode is triggered, the geological variability index value of the current sampling point is obtained. This means that once the system detects that the geological variability index exceeds the first threshold, it immediately records the specific value of the index.

[0121] Based on the correspondence between the geological variability index value and the preset threshold range, the pre-stored sampling density level is matched. This means that the system has multiple preset threshold ranges, each of which corresponds to a sampling density level.

[0122] The specific mapping table and number of sampling density levels are as follows: The system presets four sampling density levels, which are progressively increased in density according to the complexity of the geological strata. Level 0 is the initial sampling interval, corresponding to a geological variability index less than the first threshold, for example, 0.3, with a sampling interval of 2 meters. Level 1 is lightly dense, corresponding to a geological variability index between the first threshold and 0.5, with a sampling interval of 1 meter. Level 2 is moderately dense, corresponding to a geological variability index between 0.5 and 0.7, with a sampling interval of 0.5 meters. Level 3 is heavily dense, corresponding to a geological variability index greater than 0.7, with a sampling interval of 0.2 meters. The threshold range and sampling interval for each level can be adjusted according to actual needs. For example, in key exploration areas, the basic interval for level 0 can be set to 1 meter, and the density interval for level 3 can be set to 0.1 meters.

[0123] The specific process for generating the gear mapping table is as follows: First, collect the optimal sampling density corresponding to different geological variability index intervals from historical drilling data. The optimal sampling density is defined as the minimum sampling density that can completely record formation changes without missing oil and gas shows. Second, plot a scatter plot with the geological variability index as the independent variable and the optimal sampling interval as the dependent variable. A piecewise linear regression method is used to fit the mapping curve. Specifically, the data is sorted by geological variability index from smallest to largest, and dynamic programming is used to find the optimal segmentation point, minimizing the linear regression error within each segment. The number of segmentation points is preset to 3 to 5, corresponding to the number of gears. After obtaining the linear equation for each segment, the predicted value of the midpoint of each segment is taken as the sampling interval for that segment. Third, divide the gear intervals according to the mapping curve, ensuring that the optimal sampling density within each interval is basically consistent. If there is no historical data, empirical division can be used initially, such as the aforementioned 0.3, 0.5, and 0.7 segments. Subsequently, the gear threshold is adjusted once after each well is completed based on the actual drilling results.

[0124] The dynamic programming algorithm for piecewise linear regression is as follows: Given M data points, sorted by geological variability index from smallest to largest, the goal is to divide the data into K segments, where K is the preset number of segments. For each segment, fit a straight line to minimize the total error. Definition: ; .

[0125] The recurrence relation is: .

[0126] Initial state: .

[0127] After calculating the optimal segmentation point using dynamic programming, linear regression is performed on each segment to obtain the slope and intercept of each segment. The geological variability index at the midpoint of each segment is then substituted into the regression equation to obtain the predicted sampling interval for that segment, which is used as the sampling interval for that grade.

[0128] Setting the sampling interval for subsequent continuous sampling points according to the sampling density level means that the system adjusts the sampling interval of the automatic sampling device to the interval corresponding to the selected level and continues to perform sampling at that interval until the next geological variability index calculation result shows that the basic density can be restored or adjusted to another level.

[0129] The sampling density switching process is as follows: When the system calculates the geological variability index of the current sampling point and matches it to a new level, it first determines whether the new level is the same as the current level. If they are the same, the current sampling interval remains unchanged. If they are different, after completing all processing steps for the current sampling point, the system automatically adjusts the sampling interval parameters of the automatic sampling device at the start of the next sampling point. The adjustment command is sent to the programmable logic controller (PLC) via industrial Ethernet, which modifies the setting value of the sampling trigger timer or depth counter. Simultaneously, the system records the level switching event, including the switching time, depth, old and new levels, and geological variability index values, for subsequent analysis and model optimization.

[0130] The hardware system implementing this method comprises the following components: an automatic sampling device installed at the drilling fluid vibrating screen, made of corrosion-resistant stainless steel, including a robotic arm or screw conveyor, a cleaning screen, a spray system, and a dual-channel buffer mechanism. The dual-channel buffer mechanism includes two parallel sampling channels, each with an independent gripping mechanism and a cleaning screen. During normal operation, channel 1 performs the current sampling, while channel 2 stands by. When encrypted sampling is triggered, after channel 1 completes the current point, channel 2 immediately begins sampling the next point, while channel 1 performs cleaning and reset. The two channels work alternately, eliminating the gap caused by single-channel reset and achieving continuous encrypted sampling. The conveying system includes multiple corrosion-resistant conveyor belts with adjustable speed for transporting cuttings between different processing units. The multispectral rapid imaging unit is equipped with an industrial camera and LED light source; the white light source has a color temperature of 5500K, and the ultraviolet light source has a wavelength of 365 nm. The infrared drying unit has a power of 1500 watts and a temperature controllable range of 50°C to 100°C. The imaging chamber is equipped with a 20-megapixel industrial camera, a telecentric lens, a ring light source, and an ultraviolet light source. The edge computing unit utilizes industrial-grade computers equipped with GPU accelerator cards such as the NVIDIA Jetson series for real-time image processing and model inference. The edge computing unit uses the NVIDIA Jetson AGX Orin, with an AI computing power of 200 TOPS, meeting the requirements for real-time image segmentation and hybrid model inference. It has 32GB of memory and 256GB of storage, capable of caching high-definition images of the most recent 1000 sampling points. Communication with the cloud uses a 4G or 5G network; high-definition images of abnormal layers are uploaded in real time, while regular data is packaged and uploaded hourly. The control system uses a programmable logic controller to communicate with the drilling engineering monitoring system, receiving depth signals and drilling parameters, and sending sampling and equipment control commands. The specific workflow of the dual-channel buffer mechanism is as follows: During normal operation, channel 1 performs the capture and cleaning of the current sampling point, while channel 2 is in standby mode. When channel 1 completes sampling and begins cleaning and resetting, if the trigger time for the next sampling point has arrived, channel 2 is immediately started to perform sampling. If a fault occurs during the sampling process of channel 2, the system automatically switches back to channel 1 and starts it immediately after channel 1 completes cleaning, while simultaneously issuing a channel 2 fault alarm. If both channels fail simultaneously, the system will stop sampling and issue an emergency alarm, awaiting manual intervention. The cleaning time for each channel should be less than the normal sampling interval to ensure seamless transition. For encrypted sampling mode, the sampling interval may be less than the cleaning time. In this case, another channel needs to be started in advance, for example, channel 2 should be started immediately after channel 1 completes its capture to begin cleaning preparation, instead of waiting for channel 1 to finish cleaning.

[0131] In the above embodiments, by matching the geological variability index value with multiple levels, the sampling density is finely adjusted, avoiding the defects of either being too dense or not dense enough caused by simple binarization encryption, making the allocation of sampling resources more reasonable and efficient. By clarifying the specific division of sampling density levels, the generation method of the level mapping table, and the switching execution process, the high-density sampling mode has a complete logical closed loop and can operate reliably in actual drilling.

[0132] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of this application, and these improvements and substitutions should also be considered within the scope of protection of this application.

Claims

1. A cuttings logging method for intelligent sampling while drilling in oil and gas geology, characterized in that, include: Rock cuttings samples are collected in real time at a preset initial sampling interval using an automatic sampling device, and the collected rock cuttings samples are then subjected to primary cleaning. The cleaned wet rock cuttings are subjected to multi-source data acquisition, including rock cuttings image data and drilling engineering parameters; The cleaned wet rock fragments are dried to obtain dry rock fragments; Calculate the geological variability index based on the rock debris image data; A sampling control command is generated based on the comparison result between the geological variability index and a preset threshold; when the index is less than the first threshold, the current interval is maintained; when the index is greater than or equal to the first threshold, the encryption mode is switched; when the index exceeds the second threshold, an alarm is triggered and further encryption is performed. The sampling control command is sent to the automatic sampling device to dynamically adjust the sampling density of subsequent rock cuttings; When the geological variability index exceeds a preset variability threshold, a high-density sampling mode with a sampling frequency higher than the initial sampling frequency is set.

2. The cuttings logging method for intelligent sampling while drilling in oil and gas geology as described in claim 1, characterized in that, The calculation of the geological variability index includes: Extract white light images, ultraviolet light images, and real-time drilling parameters from multi-source data; Different lithological grains are segmented and identified from white light images, and the degree of lithological variation is calculated. The intensity and area of ​​fluorescent spots are extracted from ultraviolet images, and fluorescence anomaly values ​​are calculated. Obtain real-time drilling parameters corresponding to the current sampling depth and calculate the engineering anomaly index; A geological variability index is generated based on the lithological variation, fluorescence anomaly value, and engineering anomaly index.

3. The cuttings logging method for intelligent sampling while drilling in oil and gas geology as described in claim 1, characterized in that, The dynamic adjustment of the sampling density of subsequent rock cuttings includes: When the geological variability index is less than the first threshold, maintain the initial sampling interval; When the geological variability index is not less than the first threshold, the high-density sampling protocol is triggered, so that subsequent consecutive sampling points adopt a sampling mode with a smaller than the initial sampling interval. When the geological variability index exceeds the second threshold, an alarm for an abnormal stratum is simultaneously sent to the control room.

4. The cuttings logging method for intelligent sampling while drilling in oil and gas geology as described in claim 1, characterized in that, The wet rock fragments are dried to obtain dry rock fragments, and the process further includes: The cleaned wet rock fragments are sent to the drying unit for drying treatment; The dried rock fragments were spread out and sent into the imaging chamber to acquire high-definition white light and ultraviolet light images. Based on the high-resolution white light and ultraviolet light images, lithology identification and oil-bearing rating are performed to generate preliminary analysis results.

5. The cuttings logging method for intelligent sampling while drilling in oil and gas geology as described in claim 4, characterized in that, The lithological identification includes: The theoretical lag time is calculated in real time based on drilling fluid discharge and annular cross-sectional area parameters. The theoretical lag time is then corrected to obtain the actual depth offset. The acquisition depth of the cuttings image is aligned and matched with the corresponding depth of the drilling engineering parameters based on the actual depth offset.

6. The cuttings logging method for intelligent sampling while drilling in oil and gas geology as described in claim 5, characterized in that, The lithological identification and oil-bearing rating also include: Construct a dual-channel hybrid model that includes image feature extraction channels and engineering parameter feature extraction channels; Lithological feature vectors are extracted from high-resolution images using the image feature extraction channel; The engineering response feature vector is extracted through the engineering parameter feature extraction channel by using a drilling parameter sequence aligned with the depth of the cuttings image; By integrating the lithological feature vector and the engineering response feature vector, the lithological category, oil-bearing level, and corresponding confidence level are output.

7. The cuttings logging method for intelligent sampling while drilling in oil and gas geology as described in claim 6, characterized in that, The dual-channel hybrid model also includes: Obtain historical data from existing oil fields; Based on the historical data, a pre-trained image feature extraction channel is trained through self-supervised learning. For the oilfield to be explored, the fully connected classification layer of the dual-channel hybrid model is fine-tuned; The training process eliminates the differences in characteristic distribution between existing oil fields and oil fields to be explored.

8. The cuttings logging method for intelligent sampling while drilling in oil and gas geology as described in claim 4, characterized in that, It also includes collaborative learning: Obtain the automatic analysis results and store the manual correction records of the automatic analysis results in the experience playback buffer; At preset intervals, samples are extracted from the experience replay buffer and combined with historical data for training, and the dual-channel hybrid model is updated.

9. The cuttings logging method for intelligent sampling while drilling in oil and gas geology as described in claim 8, characterized in that, The fusion of the lithological feature vector and the engineering response feature vector includes: Calculate the attention weight of the lithological feature vector to the engineering response feature vector; The engineering response feature vector is reorganized based on the attention weights to generate engineering enhancement features; The lithological feature vector is concatenated with the engineering enhancement feature and input into the fully connected layer for classification.

10. The method for intelligent sampling of oil and gas geology during drilling, as described in claim 9, is characterized in that... The dynamic adjustment of the sampling density of subsequent rock cuttings includes: When the high-density sampling mode is triggered, the geological variability index value of the current sampling point is obtained; Based on the correspondence between the geological variability index value and the preset threshold range, the pre-stored sampling density level is matched; Set the sampling interval for subsequent consecutive sampling points according to the sampling density level.

Images