Fluid inclusion intelligent judgment and temperature measurement method and system based on deep learning

Abstract

The invention provides a fluid inclusion intelligent judgment and temperature measurement method and system based on deep learning, and relates to the field of mineral analysis and visual identification, and the method comprises the steps: achieving the recognition of a fluid inclusion through dynamic and static feature extraction and fusion, and recognizing a key frame of a key region in a phase change process; the foreground and the background are distinguished more accurately by introducing an attention mechanism, so that the recognition accuracy of the fluid inclusion is improved; then the temperature of the fluid inclusion in the area is output through the multiple parallel double-channel temperature measurement layers, the recognition accuracy of the fluid inclusion is improved, the temperature measurement accuracy is achieved, a multi-constraint active learning mechanism is introduced in the training process, and the training accuracy is improved. The workload of temperature marking is reduced, and training is effectively prevented from falling into a local optimal solution; meanwhile, the temperature measurement layer can output a phase change key frame picture series of a key area, and visual output is achieved.

Description

Technical Field

[0001] This invention relates to the fields of mineral analysis and visual recognition, and in particular to a method and system for intelligent judgment and temperature measurement of fluid inclusions based on deep learning. Background Technology

[0002] Fluid inclusions refer to fluids trapped within mineral lattice defects and micropores during diagenesis and mineralization. Their composition can include atmospheric precipitation, formation water, gaseous hydrocarbons, and liquid hydrocarbons. Fluids in oil and gas basins, including atmospheric precipitation, formation water, gaseous hydrocarbons, and liquid hydrocarbons, have migrated and accumulated over a long geological history, recording fluid information from different periods during diagenesis as fluid inclusions. Analysis of fluid inclusions has become an important tool for studying basin fluid behavior and hydrocarbon accumulation mechanisms, widely used in identifying oil and gas reservoirs, tracing migration paths, determining accumulation ages, analyzing accumulation conditions, and reconstructing accumulation processes. Through lithofacies observation, microthermometry, compositional analysis, and isotopic dating of fluid inclusions, research plays a crucial role in oil and gas resource evaluation, reservoir geochemistry analysis, identification of oil and gas fluid types and accumulation phases, and revelation of fluid sources. Traditional methods for studying fluid inclusions primarily rely on manual observation and measurement. First, researchers carefully observe thin sections of fluid inclusions under a microscope, manually counting the number of inclusions in the field of view. They then use micrometers to measure the size of the inclusions, roughly estimating the gas-liquid ratio to obtain a qualitative description of the inclusion's morphology. This method suffers from low efficiency and insufficient accuracy. Furthermore, observing and measuring the temperature of fluid inclusions under a polarizing microscope requires operators with extensive knowledge of fluid inclusions and proficient microscope skills. For large sample analyses, the entire process of finding, locating, and measuring inclusions is tedious, time-consuming, and inefficient. More importantly, the analytical results are easily influenced by the operator's personal experience and subjective judgment, affecting the comparability and accuracy of the experimental results. Due to limitations in the precision of measuring tools, traditional manual analysis methods are prone to misjudging or missing key phase transition moments for small or complex fluid inclusions, failing to meet the geological research needs for rapid and accurate analysis of fluid inclusion parameters. In recent years, deep learning technologies, represented by convolutional neural networks, have demonstrated superior robustness, fast processing speed, and high automation in various fields such as computer vision, thanks to their powerful feature extraction and complex image recognition capabilities. Although some studies have attempted to use image processing algorithms (such as thresholding or edge detection) to assist analysis, these methods have poor generalization ability for complex backgrounds and multi-morphic inclusions, and cannot achieve end-to-end automated processing. While deep learning-based models offer excellent performance, they typically rely on large amounts of labeled data, facing challenges in the geological field such as scarce labeled samples, cumbersome labeling processes, and low accuracy. Summary of the Invention

[0003] Based on the above problems, this invention proposes an intelligent analysis framework that integrates multimodal feature extraction, ensemble learning and active learning, to achieve full-process automation from inclusion identification and localization to automatic interpretation of phase transition temperature, significantly improving analysis efficiency and accuracy.

[0004] The first aspect of this invention provides a deep learning-based intelligent judgment and temperature measurement method for fluid inclusions. This method is implemented through a fluid inclusion intelligent judgment and temperature measurement model, which includes a contour recognition layer, a dynamic and static feature fusion layer, and a temperature measurement layer. The temperature measurement layer includes multiple parallel sub-temperature measurement layers. The method includes: Step S101, converting the input video of the fluid inclusion sheet into a high-resolution image sequence over time, and obtaining the bounding box coordinates, size, shape, and category information of the fluid inclusion through the contour recognition layer; Step S102, based on the bounding box coordinates... The image sequence is cropped into a key area image sequence based on the label, size, shape, and category information. The key area sequence is the region where the fluid inclusions are in the phase transition process. The key area image sequence is formed into a phase transition video based on the time relationship. Step 103: The static features of each key area image sequence and the dynamic features of the phase transition video are extracted by the dynamic and static feature fusion layer, and the static features and the dynamic features are fused into a fused feature. Step S104: The fused feature is analyzed by each of the sub-temperature measurement layers to obtain multiple sub-predicted temperatures. The mean of each sub-predicted temperature is determined as the predicted temperature, and the variance of each sub-predicted temperature is determined as the uncertainty.

[0005] In some embodiments, prior to step S101, the intelligent judgment and temperature measurement method for fluid inclusions further includes: step S201, inputting an unlabeled image sequence into the untrained temperature measurement layer to obtain a predicted temperature and uncertainty, wherein the initial values ​​of each sub-temperature measurement layer in the untrained temperature measurement layer are different; step S202, progressively selecting images in the image sequence with uncertainties greater than a preset uncertainty threshold for temperature labeling to obtain the actual temperature, and iteratively training the temperature measurement layer using the temperature-labeled images until the training of the temperature measurement layer meets the training completion conditions.

[0006] In some embodiments, step S202 includes: step S301, identifying the top 20% of images with the highest uncertainty as a candidate set; step S302, performing cluster analysis based on the fusion features of each image to obtain multiple clusters and identifying each cluster as multiple training sets; step S303, randomly selecting images from each cluster to form multiple training sample sets; step S304, sequentially inputting each image from each training sample set into the temperature measurement layer to obtain multiple predicted temperatures and multiple uncertainties; step S305, arranging the uncertainties from largest to smallest and identifying the images corresponding to the top preset uncertainties as a set to be labeled, and labeling the images in the set to be labeled with temperature to obtain the actual temperature; step S306, obtaining the loss value through the difference between the actual temperature and the predicted temperature, backpropagating the loss value and adjusting the weight coefficient of the temperature measurement layer; step S307, repeating steps S304 to S306 until the loss value continuously converges to an interval less than a preset threshold.

[0007] In some embodiments, between step S302 and step S303, step S202 further includes: removing images from each cluster whose distance from the cluster center point is greater than a preset distance threshold.

[0008] In some implementations, step S305 includes: calculating the mean of the predicted temperatures corresponding to all the unlabeled images to obtain the mean temperature; removing images from the set to be labeled whose difference between the predicted temperature and the mean temperature is greater than a preset temperature threshold; and labeling the images in the removed set to be labeled with temperature to obtain the actual temperature.

[0009] In some implementations, step S201 includes: inputting an unlabeled image sequence into the untrained thermometric layer to obtain predicted temperature, uncertainty, and fusion features.

[0010] In some embodiments, the contour recognition layer includes a residual layer and an attention layer, and step S101 includes: recognizing the bounding box coordinates, size, shape and category information of the fluid including the volume through the residual layer, and improving the recognition accuracy of the blurred boundary area between the foreground and the background through the attention layer.

[0011] In some embodiments, prior to step S101, the intelligent judgment and temperature measurement method for fluid inclusions further includes: collecting samples with annotation information of fluid inclusions, the annotation information including boundary position information and type information of the fluid inclusions; generating amplified training samples by generating adversarial networks and the samples with annotation information of fluid inclusions; and training the contour recognition layer and the dynamic and static feature fusion layer using the amplified training samples.

[0012] In some embodiments, after step S104, the intelligent judgment and temperature measurement method for fluid inclusions further includes: identifying an image sequence with phase transition keyframes; and fusing the predicted temperature and the image sequence to obtain an image sequence with temperature information.

[0013] A second aspect of this invention provides a deep learning-based intelligent judgment and temperature measurement system for fluid inclusions. The system includes: a contour recognition module, used to convert an input video of a fluid inclusion sheet into a set of high-resolution image sequences over time, obtaining the bounding box coordinates, size, shape, and category information of the fluid inclusion through the contour recognition layer; a feature fusion module, used to crop the high-resolution image sequences into key region image sequences through the contour recognition layer, the key region sequence being the region where the fluid inclusion is undergoing a phase transition, forming a phase transition video based on time relationships; further used to extract static features from each key region image sequence and dynamic features from the phase transition video through the dynamic and static feature fusion layer, fusing the static features and dynamic features into a fused feature; and a temperature measurement module, used to analyze the fused feature through each self-temperature measurement layer to obtain multiple sub-predicted temperatures, determining the mean of each sub-predicted temperature as the predicted temperature, and determining the variance of each sub-predicted temperature as the uncertainty.

[0014] By using a dual-channel dynamic and static feature fusion layer (3D-CNN+2DResNet) for feature extraction and adaptive fusion mechanism, the static and dynamic features of different types of encapsulations are extracted and fused to obtain fused features, and the predicted temperature based on these fused features is more accurate. Attached Figure Description

[0015] Figure 1 A flowchart illustrating a deep learning-based intelligent judgment and temperature measurement method for fluid inclusions provided in an embodiment of the present invention; Figure 2 A schematic diagram of the residual network that incorporates an attention framework; Figure 3 This is a schematic diagram of a deep learning-based intelligent judgment and temperature measurement model for fluid inclusions. Figure 4 A flowchart illustrating another method for intelligent identification and temperature measurement of fluid inclusions based on deep learning, provided in an embodiment of the present invention. Figure 5 for Figure 4 A detailed flowchart of step S202 is shown below; Figure 6 This is a schematic diagram of a deep learning-based intelligent judgment and temperature measurement system for fluid inclusions, provided in an embodiment of the present invention. Detailed Implementation

[0016] The specific technical features described in the various embodiments in the detailed implementation can be combined in various ways without contradiction. For example, different implementation methods can be formed by combining different specific technical features. In order to avoid unnecessary repetition, the various possible combinations of the specific technical features in this invention will not be described separately.

[0017] It should also be noted that, in order to avoid obscuring the present invention with unnecessary details, only the structures and / or processing steps closely related to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0018] Additionally, it should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the following description, the terms "first," "second," etc., are used merely to distinguish different objects and do not indicate any similarity or connection between them. It should be understood that the directional descriptions such as "above," "below," "inside," and "outside" refer to the orientation under normal use conditions.

[0019] In the following specific embodiments, the intelligent judgment and temperature measurement method for fluid inclusions can be applied to the identification and temperature measurement of fluid inclusions in any rock formation or mineral. The fluid inclusions can be, for example, oil and gas fluids encased in the rock formation. The steps of the method are illustrated below with reference to various embodiments.

[0020] In some embodiments, the intelligent judgment and temperature measurement method for fluid inclusions is implemented through an intelligent judgment and temperature measurement model for fluid inclusions. This model includes a contour recognition layer, a dynamic and static feature fusion layer, and a temperature measurement layer. The temperature measurement layer includes multiple sub-temperature measurement layers connected in parallel, such as... Figure 1 As shown, the intelligent judgment and temperature measurement method for fluid inclusions includes: Step S101: Convert the input video of the fluid inclusion thin slice into a set of high-resolution image sequences over time, and obtain the bounding box coordinates, size, shape and category information of the fluid inclusion through the contour recognition layer.

[0021] This can be understood as follows: First, the recorded video of fluid inclusion sections is converted into a sequence of images arranged chronologically. A feature pyramid network is then used to perform multi-scale sampling and fusion of each image to improve image resolution. Next, image recognition is used to extract the bounding box coordinates, size, shape, and category information of the fluid inclusions in the images. The fluid inclusion sections are prepared from rock formations containing fluid inclusions.

[0022] Step S102: Based on the bounding box coordinates, size, shape and category information, the image is cropped into a key area image sequence, and based on the time relationship, the key area image sequence is formed into a phase transition video.

[0023] This can be understood as follows: By using bounding box coordinates, size, shape, and category information, we determine which fluid inclusions are in a phase transition state and in which images this phase transition state is present (since these images are arranged chronologically, they also reflect the time period of the phase transition; for ease of explanation, each image is referred to as a frame). Spatially, the images are cropped to retain the fluid inclusions in phase transition and their surrounding areas, called the key region. Temporally, key frames in and around the phase transition process are retained. For example, if the phase transition process lasts 44 frames, in addition to retaining the key region images of these 44 key frames, key region images of the first 10 frames and the last 10 frames of the phase transition process are also extracted, resulting in a 64-frame key region image sequence. These key region image sequences are then stitched together chronologically to obtain the phase transition video, or the phase transition video can be directly extracted from the original video based on the key regions and time points. This process significantly reduces the amount of data processing required in subsequent steps.

[0024] Step S103: Extract the static features of the image sequence of each key area and the dynamic features of the phase transition video through the dynamic and static feature fusion layer, and fuse the static features and dynamic features into fused features.

[0025] Specifically, the dynamic and static feature fusion layer is a dual-channel network, including a 3D-CNN dynamic feature extraction channel and a 2DResNet static feature extraction channel. The 3D-CNN dynamic feature extraction channel uses 3D convolutional kernels (e.g., 3×3×3) to perform convolution operations simultaneously in both spatial and temporal dimensions, enabling the learning of motion and information changes over time. It can capture the minute tremors (Brownian motion) of bubbles before homogenization, the blurring-clarification process of phase interfaces, and the transient dynamic features at the moment of bubble disappearance to determine phase transitions and thus accurately measure temperature. Based on 2DResNet, keyframes are extracted from the video cube to focus on extracting high-resolution spatial texture features and analyzing static appearance information such as the gas-liquid ratio of inclusions, the sharpness of the two-phase interface, the transparency within the liquid phase, the presence of daughter minerals, and their crystal forms. This information provides important contextual supplementation for the dynamic features.

[0026] Optionally, based on this, an SE (Squeeze-and-Excitation) attention mechanism is introduced to enhance the perception of texture features such as gas-liquid ratio, interface sharpness, and sub-mineral morphology. The feature vectors from the two channels are fused. This fusion module automatically learns a set of weights to dynamically evaluate the importance of dynamic and static features at the current moment. The fused feature vector is a powerful representation that combines spatiotemporal and textural information, providing information support for the final temperature regression. For example, Figure 2 As shown, the SE module models the relationship between channels by introducing a Squeeze operation and an Excitation operation. In the Squeeze stage, it compresses the output feature map of the convolutional layer into a feature vector through a global average pooling operation. Then, in the Excitation stage, it learns to generate a channel weight vector by using a fully connected layer and a non-linear activation function (such as the sigmoid function). This weight vector is applied to each channel on the original feature map to weight the features of different channels. In this way, the SE module can adaptively learn the importance of each channel and adjust the channel contribution in the feature map according to the needs of the task. This attention mechanism helps the network to better focus on important feature channels, thereby improving model performance.

[0027] Step S104: Analyze the fusion features of each sub-temperature measurement layer to obtain multiple sub-predicted temperatures, determine the mean of each sub-predicted temperature as the predicted temperature, and determine the variance of each sub-predicted temperature as the uncertainty.

[0028] This can be understood as follows: by setting different parallel sub-channels to analyze the fusion features, different sub-predicted temperatures are obtained, and the average temperature is determined as the output predicted temperature, thus making the predicted temperature more accurate. Moreover, the reliability of the predicted temperature can be determined by the uncertainty.

[0029] By using a dual-channel dynamic and static feature fusion layer (3D-CNN+2DResNet) for feature extraction and adaptive fusion mechanism, the static and dynamic features of different types of encapsulations are extracted and fused to obtain fused features, and the predicted temperature based on these fused features is more accurate.

[0030] The following is combined Figure 3The fluid inclusion intelligent judgment and temperature measurement model illustrated provides an overall example of data flow and processing. First, a time-series image of fluid inclusions is input into a contour recognition layer. This layer, comprising ResNet (Residual Network) and FPN (Feature Pyramid Network), processes and recognizes the images, resulting in high-resolution images with labeled information. This labeled information includes the bounding box coordinates, size, shape, and category information of the fluid inclusions. Then, this information is used to crop the original images and videos, retaining only keyframe images of key regions before and after the phase transition, as well as video footage of the phase transition period within those key regions. To reduce the amount of data processing in subsequent processes and minimize the impact of non-critical areas on subsequent processing, the cropped images and videos are input into a dynamic and static feature fusion layer (including parallel 3D-CNN and 2DResNet layers). Dynamic features are extracted using the 3D-CNN dynamic feature extraction channel, and static features are extracted using 2DResNet. The dynamic and static features are then fused together using an attention layer to obtain the fused features. The fused features are input into each sub-temperature measurement layer in the temperature measurement layer, so that each sub-temperature measurement layer outputs a different sub-predicted temperature. The mean of each sub-predicted temperature is determined as the predicted temperature, and the variance of each sub-predicted temperature is determined as the uncertainty.

[0031] In some embodiments, such as Figure 4 As shown, with Figure 1 The fluid inclusion intelligent judgment and temperature measurement method shown is different. Figure 1 Prior to step S101, the intelligent judgment and temperature measurement method for fluid inclusions further includes: Step S201: Input the unlabeled image sequence into the untrained thermometric layer to obtain the predicted temperature and uncertainty.

[0032] In this case, the initial values ​​of each sub-temperature measurement layer in the untrained temperature measurement layer are different, so that the sub-predicted temperatures output by each sub-temperature measurement layer are different after inputting an image sequence without temperature labeling, thus obtaining different uncertainties.

[0033] Step S202: Gradually select images in the image sequence whose uncertainty is greater than a preset uncertainty threshold, and perform temperature labeling to obtain the actual temperature. Iteratively train the temperature measurement layer using the temperature-labeled images until the training of the temperature measurement layer meets the training completion conditions.

[0034] It's important to note that labeling the temperature of fluid inclusions in images requires significant effort and cost. Indiscriminately labeling a large number of samples would result in enormous training costs for the temperature measurement model. Therefore, a partially labeled training method is necessary, while also preventing training from getting trapped in local optima due to small sample sizes. Uncertainty can be used for selection here. The uncertainty table, besides representing the reliability of the prediction results, can also represent the dispersion of sub-predicted temperatures output by multiple sub-temperature measurement models during training. This indicates that the input samples have sufficient information to ensure sufficient dispersion of sub-predicted temperatures output by multiple structurally identical sub-temperature measurement layers. Gradually selecting samples with sufficiently high uncertainty for training ensures that the samples have a sufficiently large information content; that is, information content constraints reduce the possibility of premature convergence due to training getting trapped in local optima. Specifically, the method of gradually selecting samples can be to randomly select from all samples with uncertainties greater than a preset threshold and iteratively train until the training meets the completion conditions.

[0035] In some embodiments, such as Figure 5 As shown, Figure 1 Step S202 includes: Step S301: Select the images with the highest uncertainty as the candidate set.

[0036] That is, candidate samples to be labeled are initially constructed through uncertainty.

[0037] Step S302: Perform cluster analysis based on the fusion features of each image to obtain multiple clusters and determine each cluster as a multi-batch training set.

[0038] Optionally, the fusion features can be obtained through the pre-dynamic and static feature fusion layer of the temperature measurement layer. Optionally, the structure of each sub-temperature measurement module is similar to that of the dynamic and static feature fusion layer, which is a dual-channel parallel structure including a 3D-CNN layer and a 2DResNet layer. The difference is that the output layer of the fusion module is set to output sub-predicted temperature and sub-fusion features. By fusing the sub-fusion features, the fusion features of the sample can be obtained, so that the temperature measurement layer can be trained independently.

[0039] A multidimensional feature space is constructed by fusing the elements in the features. The elements in the fusing features are used as coordinates to form points in the multidimensional feature space, and cluster analysis is performed to obtain multiple clusters. Different clusters represent different sample types. The cluster analysis method can be K-cluster analysis, X-cluster analysis, or density-based cluster analysis.

[0040] Optionally, after forming each cluster, it is also necessary to remove images from each cluster whose detection distance from the cluster center point is greater than a preset distance threshold. This removes samples in each cluster that differ excessively from the rest of the samples, thereby improving the training process. Step S303: Randomly select images from each cluster to form multiple training sample sets.

[0041] This can be understood as the fact that each cluster may still contain a large number of samples, thus requiring the formation of training sample sets with different types by randomly sampling from different clusters.

[0042] Step S304: Input each image from each training sample set into the temperature measurement layer in sequence to obtain multiple sets of predicted temperatures and multiple sets of uncertainties.

[0043] This can be understood as rotating the training sample sets from different clusters into the temperature measurement layer, so that the input samples belong to different clusters during at least several generations of iterative training. That is, by inputting different types of samples, the training is further constrained by type, which further reduces the possibility of the training converging to a local optimum. It should be noted that the training sample sets are re-inputted into the temperature measurement layer here because the weight coefficients of the temperature measurement layer have changed in subsequent iterative training, so that the uncertainty obtained is different from the uncertainty obtained from the initial temperature measurement layer.

[0044] Optionally, before temperature labeling, it is also necessary to calculate the mean of the predicted temperature obtained after inputting each training set into the temperature measurement layer, and remove samples corresponding to temperatures that differ too much from the mean of that temperature. That is, remove samples with excessively large temperature differences, thereby further representing the training and enabling the model to learn only mainstream samples.

[0045] Step S305: Arrange the uncertainties from largest to smallest and determine the images corresponding to the previous preset uncertainties as the set to be labeled. Then, label the images in the set to be labeled with temperature to obtain the actual temperature.

[0046] This can be understood as further reducing the number of training samples by labeling the graphs corresponding to the top few uncertainties.

[0047] Step S306: Obtain the loss value by the difference between the actual temperature and the predicted temperature, backpropagate the loss value and adjust the weight coefficient of the temperature measuring layer.

[0048] This can be understood as constructing a function using the actual temperature and the predicted temperature to obtain the loss value, then backpropagating the loss value and adjusting the weighting coefficient of the temperature sensing layer.

[0049] Step S307: Repeat steps S304 to S306 until the loss value continuously converges to an interval less than a preset threshold.

[0050] That is, the training is completed when the loss value continuously converges to an interval less than the preset threshold, indicating that the temperature measurement layer has been able to predict the temperature stably and accurately.

[0051] In summary, by employing a triple constraint of information content, type, and representativeness, temperature labeling is progressively selected from the overall sample while avoiding the possibility of training getting trapped in local optima with small sample sizes. Specifically, as training progresses, if all clusters have been input, the sample set is drawn again in the same order. Since the probability of drawing samples that have already been temperature-labeled increases, the number of samples requiring labeling gradually decreases during training. If all drawn samples are already labeled, temperature labeling is unnecessary for this training iteration, significantly reducing the workload of temperature labeling.

[0052] In other embodiments, the temperature sensing layer can also obtain an image sequence with phase transition keyframes based on the phase transition keyframes in the image fusion features. At the same time, it can combine these image sequences with the predicted temperature to output an image sequence with temperature information, which can be expressed by, for example, by labels and temperature numbers, or by color to represent the corresponding temperature.

[0053] In some embodiments, the contour recognition layer includes a residual layer and an attention layer. Figure 1 Step S101 includes identifying the bounding box coordinates, size, shape, and category information of the fluid, including the volume, through the residual layer; and improving the recognition accuracy of blurred foreground and background boundaries through the attention layer. Specifically, RetinaNet finely adjusts its attention by incorporating Focal Loss, prioritizing details that are easily overlooked but crucial to overall recognition accuracy, thus resolving the category imbalance problem between the foreground (inclusions) and background (host minerals) during detection. The classification loss (…) ) is used to predict the target category, bounding box regression loss ( This is used for precise target location. Based on this, a Feature Pyramid Network (FPN) is introduced to fuse feature maps from different levels in the backbone network, generating a set of pyramid-shaped feature maps with multi-scale features. This model can effectively extract features and train for detection on images of various scales, thereby improving the accuracy of target detection. The network model is input with a data-augmented microscopic image, and for each detected inclusion, it outputs its bounding box coordinates, size, shape, and category information, providing analysis data for subsequent temperature measurement.

[0054]

[0055]

[0056]

[0057] In the formula: CE (crossentropy): cross-entropy loss; FL (FocalLoss): attention loss; p: estimated probability of the model, p∈(0,1); γ: adjustable focusing parameter; modulation factor: (1-pt) γ Lcls is the attention loss after processing with the sigmoid function; Lreg: L1Loss (representing the difference between the actual and predicted parameters); i: all positive and negative samples; j: all positive samples; Npos: the number of positive samples.

[0058] In some embodiments, prior to step S101, the contour recognition layer and the dynamic-static feature fusion layer need to be trained with a large number of labeled samples. This labeled information includes the boundary location and type information of the fluid inclusions. Since the number of samples may be insufficient, amplified samples need to be generated using a generative adversarial network (GAN). These amplified samples are then used to train the contour recognition layer and the dynamic-static feature fusion layer to improve the training sample size. Specifically, double-polished thin sections of inclusions prepared according to standards are collected. These sections are then observed, photographed, and recorded under a research-grade microscope equipped with a high-resolution digital microscope camera, while maintaining image clarity and detail for subsequent annotation and analysis. The microscope integrates a Linkam hot and cold stage with a temperature control accuracy better than ±0.1°C. Images with dense distributions and diverse morphologies of fluid inclusions are selected, and experienced fluid inclusion researchers are invited to guide the fluid inclusion annotation process. The LabelImg annotation tool is used to perform fine bounding box annotations on each fluid inclusion in the static microscopic image. The annotation information includes (x_min, y_min, x_max, y_max, class). The dataset can be further subdivided into categories such as (pure liquid hydrocarbon inclusions, multiphase pure liquid hydrocarbon inclusions, gaseous hydrocarbon inclusions, bitumen inclusions, and gas-liquid two-phase hydrocarbon inclusions) to ensure the sufficiency and reliability of the dataset. Based on this, a GAN (Generative Adversarial Network) is used for data augmentation and expansion. The original images of the inclusions are input into the GAN network model for training, converting low-resolution images into high-resolution images to improve image quality. Simultaneously, pseudo-images that are highly similar to real inclusions in morphology, texture, and lighting, but are entirely new, are generated, forming a high-resolution, standard-compliant training set.

[0059] This invention also provides a deep learning-based intelligent judgment and temperature measurement system for fluid inclusions, such as... Figure 6As shown, the intelligent fluid inclusion judgment and temperature measurement system includes: a contour recognition module 100, a feature fusion module 200, and a temperature measurement module 300. The contour recognition module 100 converts the input video of the fluid inclusion slice into a set of high-resolution image sequences over time, obtaining the bounding box coordinates, size, shape, and category information of the fluid inclusion through the contour recognition layer. The feature fusion module 200 crops the high-resolution image sequences into key area image sequences through the contour recognition layer. The key area sequence represents the region where the fluid inclusion is undergoing a phase transition, and the key area image sequences are combined into a phase transition video based on the time relationship. It also extracts the static features of each key area image sequence and the dynamic features of the phase transition video through a dynamic and static feature fusion layer, fusing the static and dynamic features into a fused feature. The temperature measurement module 300 analyzes the fused features through its respective temperature measurement layer to obtain multiple sub-predicted temperatures, determining the mean of each sub-predicted temperature as the predicted temperature and the variance of each sub-predicted temperature as the uncertainty.

[0060] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent identification and temperature measurement of fluid inclusions based on deep learning, characterized in that, The intelligent judgment and temperature measurement method for fluid inclusions is implemented through an intelligent judgment and temperature measurement model for fluid inclusions. This model includes a contour recognition layer, a dynamic and static feature fusion layer, and a temperature measurement layer. The temperature measurement layer comprises multiple parallel sub-temperature measurement layers. The intelligent judgment and temperature measurement method for fluid inclusions includes: Step S101: Convert the input video of the fluid inclusion thin slice into a set of high-resolution image sequences over time, and obtain the bounding box coordinates, size, shape and category information of the fluid inclusion through the contour recognition layer; Step S102: Based on the bounding box coordinates, the size, the shape and the category information, the image sequence is cropped into a key area image sequence, the key area sequence is the region where the fluid inclusions in the phase transition process are located, and the key area image sequence is formed into a phase transition video based on the time relationship; Step 103: Extract the static features of each key region image sequence and the dynamic features of the phase transition video through the static and dynamic feature fusion layer, and fuse the static features and the dynamic features into a fused feature; Step S104: Analyze the fusion features through each of the sub-temperature measurement layers to obtain multiple sub-predicted temperatures, determine the mean of each sub-predicted temperature as the predicted temperature, and determine the variance of each sub-predicted temperature as the uncertainty.

2. The intelligent judgment and temperature measurement method for fluid inclusions according to claim 1, characterized in that, Prior to step S101, the intelligent judgment and temperature measurement method for fluid inclusions further includes: Step S201: Input the unlabeled image sequence into the untrained temperature measurement layer to obtain the predicted temperature and uncertainty, wherein the initial values ​​of each sub-temperature measurement layer in the untrained temperature measurement layer are different. Step S202: Gradually select images in the image sequence whose uncertainty is greater than a preset uncertainty threshold, and perform temperature labeling to obtain the actual temperature. Iteratively train the temperature measurement layer using the temperature-labeled images until the training of the temperature measurement layer meets the training completion conditions.

3. The intelligent judgment and temperature measurement method for fluid inclusions according to claim 2, characterized in that, Step S202 includes: Step S301: Determine the images with the highest uncertainty as the candidate set; Step S302: Based on the fusion features of each image, perform cluster analysis to obtain multiple clusters and determine each cluster as a multi-batch training set; Step S303: Randomly select images from each cluster to form multiple training sample sets; Step S304: Input each image from each training sample set into the temperature measurement layer in sequence to obtain multiple sets of predicted temperatures and multiple sets of uncertainties; Step S305: Arrange the uncertainties from largest to smallest and determine the images corresponding to the previous preset uncertainties as the set to be labeled. Then, label the images in the set to be labeled with temperature to obtain the actual temperature. Step S306: Obtain the loss value by the difference between the actual temperature and the predicted temperature, backpropagate the loss value and adjust the weighting coefficient of the temperature measuring layer; Step S307: Repeat steps S304 to S306 until the loss value continuously converges to an interval less than a preset threshold.

4. The intelligent judgment and temperature measurement method for fluid inclusions according to claim 3, characterized in that, Between step S302 and step S303, step S202 further includes: Remove images from each cluster whose distance from the cluster center point is greater than a preset distance threshold.

5. The intelligent judgment and temperature measurement method for fluid inclusions according to claim 3, characterized in that, Step S305 includes: Calculate the mean of the predicted temperatures for all the unlabeled images to obtain the mean temperature; Images in the set to be labeled whose predicted temperature and mean temperature difference is greater than a preset temperature threshold are removed. The images in the removed set are then labeled with temperature to obtain the actual temperature.

6. The intelligent judgment and temperature measurement method for fluid inclusions according to claim 2, characterized in that, Step S201 includes: The image sequence without temperature labeling is input into the untrained thermometric layer to obtain the predicted temperature, uncertainty, and fusion features.

7. The intelligent judgment and temperature measurement method for fluid inclusions according to claim 1, characterized in that, The contour recognition layer includes a residual layer and an attention layer, and step S101 includes: The residual layer identifies the bounding box coordinates, size, shape, and category information of fluids, including volumes, while the attention layer improves the recognition accuracy of blurred foreground and background boundaries.

8. The intelligent judgment and temperature measurement method for fluid inclusions according to claim 1, characterized in that, Prior to step S101, the intelligent judgment and temperature measurement method for fluid inclusions further includes: Collect samples with annotation information of fluid inclusions, the annotation information including the boundary location information and type information of the fluid inclusions; Amplified training samples are generated by generating an adversarial network and the samples labeled with fluid inclusions. The contour recognition layer and the dynamic and static feature fusion layer are trained using the amplified training samples.

9. The intelligent judgment and temperature measurement method for fluid inclusions according to claim 1, characterized in that, Following step S104, the intelligent judgment and temperature measurement method for fluid inclusions further includes: Identify image sequences with phase transition keyframes; The predicted temperature and the image sequence are fused to obtain an image sequence with temperature information.

10. A deep learning-based intelligent judgment and temperature measurement system for fluid inclusions, characterized in that, The intelligent fluid inclusion detection and temperature measurement system includes: The contour recognition module is used to convert the input video of the fluid inclusion thin slice into a set of high-resolution images over time, and to obtain the bounding box coordinates, size, shape and category information of the fluid inclusion through the contour recognition layer; The feature fusion module is used to crop the high-resolution image sequence into a key region image sequence through the contour recognition layer. The key region sequence is the region where the fluid inclusions are in the phase transition process. Based on the time relationship, the key region image sequence is used to form a phase transition video. It is also used to extract the static features of each key region image sequence and the dynamic features of the phase transition video through the dynamic and static feature fusion layer, and fuse the static features and the dynamic features into a fused feature. The temperature measurement module is used to analyze the fusion features through each of the self-temperature measurement layers to obtain multiple sub-predicted temperatures, determine the mean of each sub-predicted temperature as the predicted temperature, and determine the variance of each sub-predicted temperature as the uncertainty.

Images