Carbonate reservoir fracture‐cavity system identification based on the improved YOLOv5s deep learning algorithm

In carbonate reservoirs characterized by the fracture‐cavity system as storage spaces, the drilling process is highly prone to the loss of drilling fluid. This not only affects drilling efficiency but can also lead to severe accidents, such as blowouts. Therefore, it is crucial to understand the distribution pattern of these fractures. However, the formation of carbonate rock fracture‐cavity system systems, being controlled by various factors, is difficult to precisely identify. This limitation hampers the efficient development of such types of oil and gas fields. This paper presents a case study of the M55 sub‐section carbonate gas reservoir in the Sulige gasfield, proposing an improved You Only Look Once v5s (YOLOv5s) deep learning algorithm. It utilizes enhanced training with conventional logging data to identify response characteristics of fractures in the carbonate reservoirs. And its identification results have been confirmed to be accurate by various fracture data obtained through different means, such as the core samples, cast thin section photographs, imaging logging data, and seismic attributes. This method incorporates the Ghost convolution module to replace the Conv module in the backbone network of the YOLOv5s model, and modifies the C3 module into a Ghost Bottleneck module, effectively making the model more lightweight. Additionally, a Convolutional Block Attention Module is integrated into the Neck network, enhancing the model's feature extraction capabilities. Finally, the method employs the Efficient Intersection over Union Loss function instead of the Complete Intersection over Union Loss, reducing the network's regression loss. The validation results using actual data demonstrate that this method achieves an average recognition accuracy of 87.3% for the fracture‐cavity system, which is a 3% improvement over the baseline model (YOLOv5s). This enhancement is beneficial for precisely locating the drilling fluid loss positions in carbonate reservoirs.

2][3] The formation and evolution of carbonate rocks are highly complex, typically undergoing multiple diagenetic processes and later modifications, resulting in substantial changes from their original structure. 4,5The most crucial change for oil and gas reservoirs is the development of numerous dissolution pores (i.e., vugs), fractures, and caves within the rock.7][8] The formation of these fractures and cavities is controlled by various diagenetic and tectonic activities at different times, making their distribution irregular and challenging to predict accurately. 9,10During drilling operations in carbonate oil and gas fields, encountering larger, network-like fractures or cavities can lead to drilling fluid leakage, severely impacting drilling efficiency.Such leakage not only causes economic loss but can also damage the reservoir's physical properties, reducing its porosity and permeability.Furthermore, it can lead to serious drilling accidents, like, blowouts, well collapse, and stuck drills. 11Therefore, effectively identifying the shape, scale, and distribution patterns of fractures and cavities in carbonate reservoirs is crucial. 12This directly influences the drilling efficiency and development outcomes of these types of oil and gas fields. 13,14urrently, the technologies for identifying fractures and cavities in carbonate rocks mainly include geophysical identification, formation microscanner image identification, conventional logging identification, and so forth. 15,16Among these, the most widely used is the geophysical identification method, which plays a crucial role in predicting well leaks.It is primarily based on fracture signals recorded in seismic data, and through complex data processing and inversion calculations, it aims to predict their distribution patterns.For instance, by utilizing typical seismic attributes, such as amplitude, amplitude variation rate, coherence, and ant-body, the distribution of the fracture and cavity systems can be determined with relative accuracy. 17The strength of seismic wave amplitude represents the differences in wave impedance properties of adjacent rock layers.The greater the difference in wave impedance, the larger the reflection coefficient and stronger the amplitude; conversely, the amplitude is weaker. 18,19If a system of fractures and cavities exists within the carbonate rock formations, due to the significant difference in wave impedance compared with the surrounding rock layers, relatively strong reflection characteristics are formed, distinguishing fracture reservoirs from other rocks. 20dditionally, the development of a fracture and cavity system in the rock layers reduces the propagation speed of seismic waves, causing changes in wave impedance and consequently leading to lateral amplitude variations.The similarity of the rock layers is disrupted, and the rate of change in amplitude attributes increases.This characteristic can be utilized to identify fracture reservoirs.These identification principles have been confirmed by a large body of research from scholars and are now widely used, especially yielding good application results in formations with large-scale fracture development and solution cavity development. 21,22Although seismic data can reflect a lot of information about fractures and cavities in carbonate rocks, for the same formation with complex systems of different scales, shapes, and distribution densities, relying solely on seismic data may not always obtain high identification accuracy.][25] In light of the limitations of using seismic data for fracture identification, some new methods and technologies have been proposed and applied, gradually improving the comprehensive accuracy of fracture identification. 26,27For example, formation microresistivity imaging (FMI) technology can directly obtain the structural features of formations within the wellbore and enable fine identification and characterization of complex pore structures in rocks. 28The color differences in FMI imaging can reflect geological information, such as lithology, fractures, and cavities.Highconductivity fractures appear as black sinusoidal lines on FMI images, while partially filled fractures show a relatively shallow sinusoidal curve.Dissolution cavities exhibit typical characteristics on FMI images, such as darker colors and gravel-like shapes, distributed in a dotted or bead-string pattern. 29In addition, acoustic image logs are one of the most effective methods for identifying fractures.Ultrasonic Borehole Imagers measure the amplitude and propagation time of ultrasonic waves emitted from the borehole wall and display the measured reflected wave amplitude and propagation time as an image in a 360°orientation within the borehole. 30,31Through extensive FMI imaging data, the dip angle, orientation, aperture, and permeability of fractures can be quantitatively analyzed.Using these reliable well data and specific algorithms, the distribution patterns of fractures can be predicted, providing insight into the fracture distribution between wells. 32Although FMI technology has advantages in fracture identification, such as large information content, high resolution, and intuitive imaging, there are still challenges in its practical application.These challenges include formation adaptability, multiple interpretations, and economic cost issues, which prevent its widespread application in all types of oil and gas reservoirs. 33,34n recent years, identifying fractures in carbonate rocks using conventional logging data has gained increasing attention from scholars.Attempts have been made in different regions and types of reservoirs, yielding favorable results. 35,36Due to the large interval between sampling points of conventional logging curves (typically 12.5 cm), the identification accuracy of a single curve is often low, making it challenging to achieve good identification results.However, combining multiple logging curves that reflect changes in rock structure and finding their different response characteristics at the same location to identify fractures is recognized as an effective method by scholars. 37,38Generally speaking, conventional logging curves that reflect lithologic characteristics (such as natural gamma, spontaneous potential, etc.) are less effective in identifying fractures. 39However, some logging curves that reflect rock properties and fluid attributes do respond to fractures and cavities.For example, the deep/shallow investigate double lateral resistivity (RLLD/RLLS) tends to show a decrease in fracture-developed layers, with the two curves presenting a positive difference, 40,41 namely, at locations with fractures and cavities, a typical "dual track" phenomenon will appear. 42Acoustic log (AC) will show an increase in transit time values in low-angle fracture development sections, and in the formation with dissolution cavities, it may even exhibit wave skipping phenomena. 43,44ompensated neutron log (CNL) studies formation properties through the interaction between neutron rays and formations.When a fracture-cavity system develops in the formation, the porosity value in the neutron log will increase significantly.When a fracture-cavity system develops in the formation, density log (DEN) will decrease significantly because the density instrument detector is in contact with open fractures. 45,46These differences in the logging responses help in identifying the presence of fractures and cavities in carbonate rock formations.In summary, this paper uses the conventional logging curves of CNL, DEN, AC, RLLD, and RLLS combinations to identify comprehensive characteristics to identify fracture-cavity development zones.The common approach for extracting fracture response from these combinations of logging curves involves a series of computational treatments of the curves, including filtering, correction, and normalization.Signal recognition is then conducted for both individual and multiple curves, followed by the use of methods like cluster analysis and principal component analysis to extract fracture information. 47Finally, the approach involves continuous verification and calibration to enhance the accuracy of the results.However, this method often relies on advanced interpretation techniques and experienced geologists, and the results may vary depending on the lithology and fluid type, affecting the precision of fracture identification.This study proposes an improved You Only Look Once version 5 (YOLOv5) method for the identification of fractures and cavities in carbonate rocks.It is a convolutional neural network-based object detection algorithm known for its fast operation speed, high detection accuracy, and real-time detection capabilities.Moreover, to achieve a lightweight yet high-precision object detection algorithm, the original YOLOv5 model is modified, using YOLOv5s as the base model for improvement.This improved model is employed to extract fracture information from a combination of five logging curves: CNL, DEN, AC, RLLD, and RLLS.The model established by this method is characterized by its high flexibility, fast detection speed, small size, low deployment cost, and strong applicability. 480][51] To enhance channel information and improve the model's ability to extract target feature information, a lightweight Convolutional Block Attention Module (CBAM) is introduced. 52Finally, the Efficient Intersection over Union (EIOU) loss function is used instead of the original Complete Intersection over Union (CIOU) loss function to accelerate convergence and enhance regression precision. 53,54| METHODS

| YOLOv5 network architecture
6][57] However, improvements and upgrades have been made to its internal organizational framework.The enhanced model features high flexibility, rapid detection speed, small size, low deployment cost, and strong applicability in practical applications (Figure 1).
Many parts of the model have been improved: First, in the Backbone section, Ghost Net replaces cross-stage partial (CSP) and CONV, achieving a lightweight YOLOv5 network and accelerating target recognition speed.Next, the attention module CBAM is added to the Neck network, enhancing the detection accuracy of features. 58The model also adopts an feature pyramid network + path aggregation network (FPN + PAN) structure (Figure 2) to improve feature diversity and facilitate the transfer of semantic information from deep to shallow feature maps. 59Finally, in the Prediction phase, the EIOU loss function replaces the original CIOU FENG ET AL.
| 2645 loss function in the network, and the Head is used to output detection results for targets of different sizes.

| Ghost model
Traditional feature extraction methods involve using a large number of convolutional kernels to extract features from images.However, stacking many convolutional kernels can lead to information redundancy, enormous computational load, and wastage (Figure 3A).In contrast, the Ghost convolution module uses fewer parameters to generate many feature maps through a series of cost-effective linear transformations, effectively revealing the information behind the inherent features.It can also be used as a plug-and-play component in combination with YOLOv5, significantly reducing the number of convolutions and parameters, thereby achieving model lightweighting and accelerating the inference process of the original network (Figure 3B). 60,61he Ghost bottleneck primarily consists of two stacked Ghost modules.When the stride is set to 1, the first Ghost module acts as an expansion layer, increasing the number of channels.The second Ghost module then reduces the number of channels to match the input channel count (Figure 4A).When the stride is set to 2, a depthwise convolution with a stride of 2 is used between the two Ghost Modules to complete the convolution scanning with a stride of 2 (Figure 4B).The CBAM is a lightweight convolutional attention module comprising two complementary modules: the CAM and the SAM (Figure 5).It can suppress redundant features and highlight effective features.By embedding it into the YOLOv5s network, not only can it save parameters and computational resources, but it can also significantly enhance detection accuracy. 62,631)

| Loss function
Among them, w c and h c are the width and height of the minimum bounding rectangle between the predicted bounding box and the actual bounding box.It is the Euclidean distance between two points.
Although CIOU_Loss considers the overlap area, center point distance, and aspect ratio of the bounding box regression, it overlooks the actual differences in width, height, and their respective confidence values, which may hinder the effectiveness of model optimization.Therefore, the EIOU_Loss function (Equation 2) is introduced, which consists of three parts: overlap loss, center distance loss, and width-height loss.(2) The first two parts continue the method in CIOU, but the aspect ratio loss item is split into the difference between the predicted value and the minimum bounding box, which accelerates convergence and improves regression accuracy (Figure 6).south (Figure 7A), and has proven reserves of more than one trillion cubic meters, making it by far the largest natural gas field in China.The gas-bearing layer is in the Ma5 section of the Majiagou Formation of the Ordovician series, which is divided from top to bottom into six sections, with the Ma5 section further divided into 10 subsections.The rock types are characterized mainly by dolomite and limestone.The reservoir depth ranges from 3100 to 3160 m, with a gentle structural relief in the region.Natural gas is primarily concentrated in potatoshaped dolomite (Figure 7C).Since dolomite in the research area is formed by later-stage metasomatism from limestone, it contains numerous fractures and cavities, making it the main area of the gas gathering.The surrounding dense micritic limestone acts as an effective seal for the fluids in the reservoir, preventing their escape (Figure 7D).By the end of December 2022, gas production from the Sourig field had exceeded 30 billion m 3 , accounting for about 58% of the total domestic onshore tight gas production.At the same time, the daily gas production of the Sourig field reached 105 million m 3 , making it a whole gas field with daily gas production exceeding 100 million m 3 . 64

| Data sets
The data used in this study consist of well logging data sets from 80 wells within the reservoir, including DEN, AC, CNL, RLLD, and RLLS logs.The LabelImg module was used to annotate the feature responses on the logging curves, resulting in 1000 images that combine these five types of log curves.The annotations include the coordinates of two diagonal points of the rectangular boxes, as well as data reflecting the location and size of the selected fractures.The data set was divided into a training set, validation set, and test set in an 8:1:1 ratio, as detailed in Table 1.

| Training parameters
In this study, the YOLOv5 algorithm's parameters are set as follows, crucial for providing an optimal environment for the algorithm to learn and optimize effectively.Initial learning rate: This key parameter influences the speed and quality of model learning.Optimizer: The optimizer used for updating and calculating the network parameters of the model.In this experiment, the default is Stochastic Gradient Descent.Workers: Number of working threads, dependent on the performance of the experimental platform.In this experiment, eight threads are chosen.These parameters ensure that the YOLOv5 algorithm is trained under optimal conditions, thereby enhancing the model's performance and accuracy (Table 2).

| Evaluation indicators
To assess the performance of the improved YOLOv5s model, this study employs a range of evaluation indicators those are crucial for understanding and comparing the performance of different models, providing a quantitative means to measure the effectiveness of a model on a specific task.The selected metrics include Box_loss, Objectness_loss, Precision (P), Recall (R), Average Precision (AP), and Mean Average Precision (mAP).The specific formulas for these metrics are detailed in the research documentation.These indicators comprehensively reflect the model's performance in object detection tasks, including its accuracy, reliability, and generalization capabilities.
Precision, also known as precision in the formula, represents the ratio of correctly predicted results among all positive samples.TP refers to the number of correctly predicted and positive samples, and FP refers to the number of incorrectly predicted and positive samples.
In the formula, Recall, also known as recall, represents the ratio of correctly predicted positive samples among all positive samples, and FN refers to the number of positive samples incorrectly predicted as negative samples.
In the formula, AP represents the average accuracy under different recall rates.
In the formula, mAP is the result of taking the mean of AP, which is the average accuracy of all categories, and can directly reflect the performance of the model.The larger the mAP value, the better the model performance and effect.The changes in various evaluation parameters during the training process are shown in Figure 8.
After 300 rounds of training, the precision (P) is around 87%, indicating that the detection accuracy is quite high.The recall rate (R) is around 83%, showing that the accuracy of finding the correct targets is reliable.The Box_loss is below 0.02, indicating precise localization, and the Obj_loss is around 0.011, suggesting high accuracy in target detection.The mAP@0.5 and mAP@0.5:0.95 are approximately 87% and 66%, respectively.From the above figure, it is evident that the performance of the improved model in terms of mAP surpasses that of the other models.Compared with Faster-RCNN, the data set used by YOLOv5s is easy to obtain, large in volume, authentic, and reliable, consisting of conventional well logging curve data.This data set ensures high accuracy and practicality.
Moreover, Faster-RCNN employs a two-stage detection method that requires two separate networks, making it relatively time-consuming.In contrast, YOLOv5 can directly predict the category and location of objects, requiring only a single feed-forward calculation for the entire process, resulting in a 3.5% increase in accuracy. 65The Single Shot MultiBox Detector (SSD) network structure is primarily designed based on a feature pyramid and multiscale prediction boxes, which makes its processing speed relatively slower.In comparison, the improved YO-LOv5s model demonstrates a 7.6% improvement in accuracy.The YOLOv5 series is an enhanced and upgraded model based on YOLOv4, boasting advantages, such as fast speed, high accuracy, and a small model size. 66Overall, among the commonly used models, the improved YOLOv5s exhibits the highest mAP value of 87.3%, surpassing Faster-RCNN, SSD, YOLOv4, and YOLOv5 models by 3.5%, 7.6%, 4.7%, and 3%, respectively.Therefore, compared with other traditional object detection algorithms, the proposed improved YOLOv5s demonstrates superior performance in terms of detection speed, accuracy, data set sources, and model deployment, making it highly feasible (Table 3).

| Identification results
The improved YOLOv5s model was used to train the established sample database, resulting in the generation of a weight file "best.pt"that reflects the fracture information of carbonate rocks.This file was then used to produce fracture probability curves.By manually comparing the identification results with the confirmed sample data, the fracture discrimination baseline was determined to be 0.5.This means that when the value of the fracture probability curve exceeds 0.5, there is a very high probability of fractures and cavities in that well section.Conversely, if the value of the fracture probability curve is less than 0.5, the likelihood of the presence of fractures and cavities in that section of the well is very low (Figure 10).Therefore, the fracture baseline and fracture probability curve can be used to identify the extent of fracture-cavity system development in each well within the target layer.

| Results verification
To conduct a comprehensive prediction of fracture and cavity system density in the entire gas reservoir, it is imperative to obtain single-well fracture data with sufficient accuracy.Therefore, it is necessary to verify the accuracy of the results identified using the YOLOv5s method to ensure their credibility.This task must be accomplished by utilizing the amount of actual well data from gas reservoirs and quantitatively characterizing the accuracy of each well.
First, fractures developed in the reservoir are identified using core samples, polarized light microscope photographs, image logs, seismic attributes, and dynamic data.The fracture parameters obtained from these data sources are relatively accurate, especially the first three, which provide precise fracture data.Subsequently, the results identified using the YOLOv5s method are compared with the actual data identification results.If the results are credible, the model can be applied to predict fractures in all wells; if the credibility is not high, the model parameters must be readjusted until the accuracy requirements are met.Finally, using a reliable model, fracture-cavity system predictions for other wells are carried out, obtaining parameters that describe the fractures (Figure 11).
(1) Core sample fracture-cavity system identification verification In the study area, there are six core sampling wells, yielding 119 m of reservoir rock samples.These data were used to identify fractures, resulting in the identification of 267 fractures and cavities, most of which are highangle calcite-filled fractures that are visually observable (Table 4).Simultaneously, the YOLOv5s model was used to identify fractures and cavities in the core segments of these wells.The results show that there were 245 sections exceeding the baseline, most of which correspond to the fractures and cavities in the core samples.That is, for the sections identified with fractures and cavities using well data, the fracture probability values were all greater than 0.5.The overall identification accuracy was 78%, meeting the accuracy requirements for predicting fracture-cavity system distribution.
(2) Microscopic fracture identification verification Due to the limitations of core sample data, only fractures and cavities with obvious external characteristics can be identified.Some may exist inside the core or be too small to be detected by the naked eye, leading to errors.Cast thin sections allow for a clearer observation of fracture   | 2653 and cavity characteristics, even those of small size, and the filling status of the fractures and cavities can also be clearly seen.In this study, 52 core cast thin sections were prepared, and 35 of these sections, observed under a microscope, showed the presence of fractures (Figure 12).Since the formation of fractures and cavities in carbonate rocks is closely related to tectonic action and fluid dissolution, they often appear in groups within the rock, rather than in isolation.This means that if fracture or cavity information is identified inside the core, there is a high probability that other fractures or cavities exist nearby.Well logging curves usually have a certain sensitivity to microscopic fractures, but this depends on factors such as the type, size, and direction of the fractures.Comparing the microscopic identification of fractures and cavities with the results from the YOLOv5s model, it is observed that areas, where fractures and cavities are visible under the microscope, can also be identified in the model.The fracture development probability curves in these areas exceed the discrimination baseline, indicating that the results obtained from both methods have a high degree of concordance (Figure 13). (

3) Imaging log data verification
Imaging logging is collecting data through 360°imaging scans of the wellbore using electrical signals.These data enable the intuitive observation of the location and orientation of fractures and cavities, and provide continuous fracture-cavity system distribution data, which is helpful in characterizing the vertical distribution pattern of fractures and cavities.The imaging log interpretation results for well SE38-60A show that numerous sinusoidal curves appeared in the well section between 3180 and 3196 m, indicating the presence of numerous high-angle fractures in a network distribution.The drilling process records for this well also reveal severe drilling fluid loss at a depth of 3183 m, suggesting a densely developed fracture-cavity system network of significant size in that section.The YOLOv5s model's identification results similarly conclude that this well section has a high density of fractures and cavities, with 80% of the fracture probability curves exceeding the discrimination baseline.This indicates that the fracturecavity system network in this region is densely clustered, confirming the credibility of the new model's identification results as verified through imaging log data (Figure 14).F I G U R E 13 Fracture-cavity system identification results and microscopic verification (SE39-62C1).
FENG ET AL.
| 2655 Analyzing the drilling fluid loss statistics in the study area (Table 5), it can be obtained that most wells experienced drilling fluid loss when drilling into the target layer (Ma5 section), just the degree of loss varied significantly.The wells located in the center of the gas reservoir experienced the most severe drilling fluid loss, while those at the edges of the gas reservoir had somewhat less severe drilling fluid loss.This indicates that the fracture-cavity system density decreases towards the periphery, and the reservoir's physical properties deteriorate, which aligns with geological understandings.The fracture probabilities for these fluid loss segments, as interpreted by the YOLOv5s model, were relatively high, all exceeding the 0.5 discrimination baseline.This suggests that the model can accurately predict the locations in the carbonate reservoir that are prone to drilling fluid loss, thus enhancing drilling efficiency.
( of the study area, specifically in the core region of the gas reservoir where the high-density wells are located.The fracture density near wells SE38-60A, SE-ZC1, and SE-ZC6 is relatively high and exhibits a network distribution.The identification results of the YOLOv5s model highly coincide with the seismic identification results.The fracture probability is relatively high in the core area surrounding well SE38-60A, where significant drilling fluid loss was also occurred during drilling; conversely, the fracture probability around well SE39-62C3 on the periphery is relatively lower, with less severe drilling fluid loss (Figure 15).This indicates that the fracture-cavity system information obtained from seismic attributes is consistent with the results identified by the new model, and the credibility of the YOLOv5s model is high.

F
I G U R E 1 Improved YOLOv5s network structure diagram.CBAM, Convolutional Block Attention Module; Conv, convolution; SPP, spatial pyramid pooling; CBL, conv and batch normalization and Leaky_relu; YOLOv5s, You Only Look Once v5s.
Bounding box regression prediction is one of the main tasks in object detection.YOLOv5 uses the CIOU_Loss function to calculate the localization loss, as shown in Equation (1).

F
I G U R E 2 FPN + PAN Structure.FPN + PAN, feature pyramid network + path aggregation network.F I G U R E 3 Ghost convolution.(A) The convolutional layer and (B) the Ghost module.F I G U R E 4 Improved Ghost bottleneck.(A) Stride = 1 bottleneck and (B) Stride = 2 bottleneck.BN, bottleneck; Conv, convolution; DW, depthwise; ReLU, rectified linear unit.IMPROVED MODEL 3.1 | Overview of the research area To validate the accuracy of the fracture-cavity system logging identification model of carbonate rocks, this study undertook applied research on the SE39-61 gas reservoir, which is located in the northern part of China's Ordos Basin.This reservoir is adjacent to the Central Paleouplift on the east and the Jingbian gas field to the F I G U R E 5 (A) The structure of the channel attention module.(B) The structure of the spatial attention module.(C) The structure of the CBAM attention module.MLP, multi-layer perceptron.F I G U R E 6 Loss function.

F
I G U R E 7 Geological overview map of the research area.(A) Tectonic units of the Ordos Basin.(B) Stratigraphic column of the Majiagou Formation in the research area.(C) Well location map of research area.(D) Gas reservoir cross-section.(according to Changqing Oilfield).
Epochs: Number of training rounds.In this process, one epoch equates to one complete training cycle using all samples in the training set.Batch-size: The batch size indicates the number of samples processed in each training cycle.Here, 16 samples are selected for training in each batch.Img-size: The pixel size of the images in the training and test sets, standardized at 640 × 640 pixels.

4 | RESULTS AND DISCUSSION 4 . 1 |
Algorithm feasibility analysisTo further validate the accuracy of the model presented in this paper, the improved model was compared with the original YOLOv5s and other mainstream object detection algorithms, such as Faster Region-based Convolutional Neural Network (Faster-RCNN), Single Shot Detector (SSD), and YOLOv4.For a fair comparison, the same training methods and parameter settings were used on the same custom data set.The results of this comparative analysis are shown in Figure9.

F
I G U R E 10 Fracture-cavity system identification results of well SE39-62C1.AC, acoustic log; CNL, compensated neutron log; DEN, density log; RLLD, deep investigate double lateral resistivity; RLLS, shallow investigate double lateral resistivity.F I G U R E 11 Model accuracy verification method.FMI, formation microresistivity imaging; YOLOv5s, You Only Look Once v5s.

( 4 )
Drilling fluid loss data verificationGenerally, if the scale of fractures in the formation is large, numerous, and interconnected into a network, severe drilling fluid loss can occur when drilling bit reaches such an area.This phenomenon can be used to assess the extent of fracture development in the reservoir.

)
Seismic attribute results verification Using ant-tracking technology, fracture attributes were extracted from the 3D seismic data in the study area.Analysis of this data indicates that the main fracture development zones are concentrated in the central part F I G U R E 14 Comparison of SE38-60A fracture-cavity system identification and prediction results.AC, acoustic log; CNL, compensated neutron log; DEN, density log; RLLD, deep investigate double lateral resistivity; RLLS, shallow investigate double lateral resistivity.
Data set division.
T A B L E 1 Comparison of results from different models.
T A B L E 4 Verification of fractures in core identification.
F I G U R E 12 Microscopic crack identification process.
T A B L E 5 Statistics of drilling fluid loss.
F I G U R E 15 Verification of seismic attribute recognition results.AC, acoustic log; CNL, compensated neutron log; DEN, density log; RLLD, deep investigate double lateral resistivity; RLLS, shallow investigate double lateral resistivity.
1) This paper proposes an improved lightweight YO-LOv5s object detection model, which reduces computational complexity, and enhances the model's regression localization capability, achieving a 3% higher accuracy than the original model.The new model can maximally extract fracture information from multiple logging curve combinations and precisely identify fracture-cavity systems in carbonate rocks.(2) The fracture-cavity system identification accuracy of the established YOLOv5s model was tested using core samples, cast thin sections, seismic attributes, imaging logs, and dynamic drilling data.The results indicate that the model's identification outcomes are highly consistent with the fracture and information obtained from actual data statistics.The model is highly credible and can be used for fracture-cavity system prediction in other wells.(3) Compared with other machine learning methods, this model has the advantage of being lightweight and fast, and has high identification accuracy.It can also be effectively applied to other fractured carbonate reservoirs.The disadvantage of this model is that it cannot distinguish between natural fractures and induced fractures, and the early training data requires a lot of processing work on five specific well log curves.