Fluid inclusion identification method based on convolutional neural network

Abstract

The invention discloses a fluid inclusion recognition method based on a convolutional neural network, and belongs to the field of oil and gas geological exploration, and the method comprises the following steps: S1, obtaining a target image of a fluid inclusion; s2, establishing a fluid inclusion image data set according to the fluid inclusion target image; and S3, according to the fluid inclusion image data set, establishing a convolutional neural network model, identifying and screening the fluid inclusion image data set, further eliminating personnel influence factors through the convolutional neural network model, improving the universality of fluid inclusion lithofacies research, and improving the accuracy of fluid inclusion lithofacies research. Key gender discrimination marks such as fluid inclusion identification, phase state types, distribution characteristics, symbiotic combination types and the like can be standardized in a unified manner, and discrimination errors caused by human factors are effectively reduced.

Description

Technical Field

[0001] The present invention relates to the field of oil and gas geological exploration, and in particular to a fluid inclusion recognition method based on convolutional neural network. Background Art

[0002] Fluid inclusions are formed when minerals capture rock-forming and ore-forming fluids during crystallization. After the formation of the inclusions, there is no internal material spilling out and no external material invading. It becomes an independent closed system and is preserved to this day. It is the only original ore-forming fluid sample. So far, the primary fluid inclusions preserved in minerals record the most original, complete and direct ore-forming environment of the ore-forming fluids. Therefore, the study of fluid inclusions has a wide range of applications in the fields of petroleum exploration, mineral deposits, etc.

[0003] The relevant research on the petrographic characteristics of fluid inclusions is the basis of all systematic research on fluids, and the identification of fluid inclusions is even more basic. At present, the identification of fluid inclusions mainly relies on the careful observation of researchers under a microscope, and further judges and identifies fluid inclusions based on previous research experience, and then extracts various key information of fluid inclusions (such as size, phase, type, distribution and symbiotic combination, etc.). This method requires a lot of time and manpower, and is prone to manual omissions and errors, poor identification quality and low reliability. Summary of the invention

[0004] In response to the above problems, the present invention provides a fluid inclusion identification method based on convolutional neural network, which can eliminate artificial factors, improve the quality of fluid inclusion identification, and reduce identification errors.

[0005] The technical solution of the present invention is:

[0006] A fluid inclusion identification method based on convolutional neural network comprises the following steps:

[0007] S1, obtaining a target image of a fluid inclusion;

[0008] S2, establishing a fluid inclusion image data set according to the fluid inclusion target image;

[0009] S3. Based on the fluid inclusion image dataset, a convolutional neural network model is established to identify and screen the fluid inclusion image dataset.

[0010] In the step S1, the target image of the fluid inclusion is acquired by using a surface scanning mirror platform, and the spatial position information of the image is recorded simultaneously.

[0011] In step S2, the acquired fluid inclusion target image needs to be pixel processed, labeled and data enhanced.

[0012] The data enhancement includes operations such as random scaling, random rotation, random horizontal flipping, Gaussian blur, and color gamut transformation on the image.

[0013] Before establishing the convolutional neural network model, the image data set needs to be divided into a training set, a validation set, and a test set.

[0014] In step S3, the fluid inclusion image data set is processed through an input layer, a convolution layer, a pooling layer, a normalization layer, an activation function, a fully connected layer and an output layer to obtain a convolutional neural network model.

[0015] The convolution layer processing method is:

[0016] C=N

[0017]

[0018] Where: C in The input feature channel dimension of the convolution layer, S is the step size, P is the padding width, H is out is the feature height of the convolutional layer output, W out H is the output feature width of the convolutional layer. in Input feature height for the convolutional layer, W in Input feature width for the convolution layer, N is the number of convolutions.

[0019] The method also includes identification verification.

[0020] The verification algorithm is as follows:

[0021]

[0022]

[0023] Among them, A represents the pixel area set of the true label, and B represents the pixel area set of the predicted value.

[0024] The beneficial effects of the present invention are:

[0025] Through the convolutional neural network model, we can further eliminate the influence of human factors, improve the universality of fluid inclusion petrographic research, and unify the key gender discrimination indicators such as fluid inclusion identification, phase type, distribution characteristics, and symbiotic combination type, effectively reducing the discrimination errors caused by human factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flow chart of a fluid inclusion identification method based on a convolutional neural network according to an embodiment of the present invention;

[0027] Figure 2It is a U-Net network structure diagram of a fluid inclusion identification method based on a convolutional neural network according to an embodiment of the present invention;

[0028] Figure 3 It is a confusion matrix diagram of a fluid inclusion identification method based on a convolutional neural network described in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0030] Example:

[0031] like Figure 1 As shown, a fluid inclusion identification method based on convolutional neural network includes the following steps:

[0032] S1, obtaining a target image of a fluid inclusion;

[0033] Digital images of fluid inclusions are collected based on the surface scanning microscope DM4P platform, and the spatial position information of these images is recorded. Compared with traditional petrographic observations of fluid inclusions, the collection efficiency and collection area are significantly improved through digital image collection. The implementation of this technology can achieve "full collection" of the information contained in the inclusion slice to be tested, avoiding omissions that are easy to occur due to manual observation.

[0034] S2, establishing a fluid inclusion image data set according to the fluid inclusion target image;

[0035] Based on the criteria of regular fluid inclusion morphology, clear inclusion outline and inclusion size ≥3μm, 870 typical fluid inclusion microscopic photos were selected as samples to construct a data set for fluid inclusion intelligent identification research.

[0036] First, the fluid inclusion microscopic images are standardized (the image resolution is set to 512×512 pixels, etc.), and then the labelme image annotation software tool based on the Pytorch deep learning framework is used to manually interpret and annotate the fluid inclusion bodies in the relevant images to achieve binary classification of inclusions and non-inclusions (including image background). The specific annotation process mainly uses the labelme annotation tool to manually interpret the inclusions in the image, circle the boundaries of the inclusions, and label the inclusion bodies with a surface label, so that the computer can distinguish the inclusions from other minerals in the image in subsequent training and learning;

[0037] Secondly, OpenCV related functions are used to perform data enhancement operations such as random scaling, random rotation, random horizontal flipping, Gaussian blur, and color gamut transformation on the inclusion image samples in the dataset, expanding the training set images from the original 870 to 36,000. Data enhancement related operations effectively expand the training samples, improve the robustness of the training model, prevent the occurrence of overfitting of the network model during training, and also allow the model to learn some invariant features, thereby improving the performance of the model.

[0038] S3. According to the fluid inclusion image data set, a convolutional neural network model is established to identify and screen the fluid inclusion image data set;

[0039] First, the data set after data enhancement operation is divided into training set, validation set and test set. Based on python language and pytorch deep learning framework, a U-Net convolutional neural network model for inclusion intelligent recognition is built. The convolutional neural network model is trained, adjusted, tested and optimized with the inclusion data set produced in the early stage as the data source, and finally the optimal convolutional neural network model for inclusion intelligent recognition application is obtained;

[0040] The convolutional neural network (CNN) of this embodiment consists of six parts: input layer, convolution layer, pooling layer, normalization layer, activation function, fully connected layer, and output layer.

[0041] The input layer is used to receive the image data input into the network. This input data contains a pixel matrix of three-dimensional information of image height, width and number of channels. In this embodiment, RGB true color inclusion image is used, and the height and width of the input pixel matrix are set to 512, and the number of channels is 3.

[0042] The input feature matrix channel dimension of the convolutional layer is C in , height and width are H in ×W in After N convolution operations with a size of K×K, a step size of S, and a padding width of P, the size of the output feature is H out ×W out ×C out , the following formula can be used to calculate the output feature size of the input feature matrix after the convolution operation.

[0043] C in =N

[0044]

[0045] The pooling layer can be regarded as a non-trainable convolutional layer. The specific hyperparameters include step size, pooling kernel size, padding, etc. The dimension of the input feature map in this embodiment is H in×W in ×C in , the pooling kernel size is F×F, and the step size is S. Then the feature dimension after pooling is:

[0046]

[0047] C out =C in

[0048] The standardization layer of this embodiment uses a batch normalization layer (Batch Normalization, BN) to maintain the feature distribution of each layer of the network so that it does not change with the iterative update of the network parameters. It standardizes the input of each layer to ensure that the data is in the sensitive range of the network activation function, significantly accelerates the training convergence speed of the network and reduces the dependence on good initialization. At the same time, by reducing the internal covariate shift, it not only enhances the generalization ability of the model and effectively alleviates the overfitting problem, but also helps control the scale of the gradient by stabilizing the activation input, combating the vanishing and explosion phenomena of the gradient, making the convolutional neural network training more stable. In addition, the implementation of BN allows the use of higher learning rates, simplifies the hyperparameter tuning process, and generally improves the training efficiency and performance of the network.

[0049] The activation function uses the ReLU function. Compared with the traditional neural network activation function, the ReLU activation function is more efficient in gradient descent and back propagation, avoiding the problems of gradient explosion and gradient disappearance; at the same time, it simplifies the calculation process, reducing the overall calculation cost of the neural network.

[0050] This embodiment does not use the traditional fully connected layer, but always maintains the spatial dimension information of the feature based on operations such as convolution, upsampling, and feature fusion to achieve pixel-by-pixel semantic segmentation tasks. The 1×1 convolution layer is used to complete the final classification-related operations and feature processing and integration in the network.

[0051] The output layer is located at the end of the convolutional neural network, following the convolutional layer, pooling layer, and fully connected layer. In the network construction of the convolutional neural network, the design of the output layer is based on the specific task of the network.

[0052] The training model used in this embodiment is carried out on the Ubuntu 20.04 platform, using Python as the main programming language, and all network models are trained based on the pytorch deep learning framework.

[0053] The digital image dataset of fluid inclusions obtained after data enhancement is input into the U-Net semantic segmentation convolutional neural network for network model training. The U-Net network model structure is as follows: Figure 2As shown in Figure 1, the U-Net network adopts the classic encoder-decoder structure.

[0054] Before training the network model, some hyperparameters in the training process need to be initialized and set. These hyperparameters can control the learning behavior, structure and optimization process of the network model. The hyperparameter configuration used in this embodiment is as shown in Table 1 below:

[0055] Table 1

[0056]

[0057]

[0058] Among them, the random seed number Seed is used to fix the random seed, so that each independent training can obtain the same result.

[0059] During the training process, it is necessary to observe the training effect and degree of the network model by observing the loss function of the network model training. Therefore, before training, it is also necessary to set the corresponding loss function to guide the learning behavior of the model during the training process.

[0060] This embodiment uses a hybrid loss function to supervise network training and optimize gradient descent. The overall loss L can be calculated according to the following formula:

[0061] L=αL BCE +βL dice

[0062] Among them, α and β represent the balance factors between the two loss functions, which are set to 0.5 and 0.5 respectively in this technology. Since this technology research is aimed at two categories, inclusions and non-inclusions, the experiment only needs to distinguish between inclusions and non-inclusions, so the loss function suitable for the binary classification model is chosen - Binary Cross-Entropy Loss (BCELoss), which can be expressed as the following formula:

[0063]

[0064] Among them, y i is the true label of sample i. When it is classified as a positive example (inclusion), the value is 1, and when it is classified as a negative example (non-inclusion), the value is 0. i It represents the probability of predicting the classification as a positive example.

[0065] During the training process, in order to better evaluate the performance of the network model, understand the effectiveness and feasibility of the network model in this technical task, and facilitate subsequent model deployment and application, it is necessary to verify the training and learning of the network model. The verification process mainly allows the network model to predict the verification set in the data set, and sets evaluation indicators to evaluate the verification effect of the network model. The verification process is often carried out after each round of training.

[0066] The data sets used in this embodiment are all binary labels, including only two categories: inclusions and non-inclusions, that is, the positive class (Positive) is the inclusion, and the negative class (Negative) is the non-inclusion. Therefore, according to the relationship between the predicted value and the true value, the prediction results can be divided into four categories:

[0067] (1) The model prediction value is positive, and the actual value is positive, which is true positive (TP, True Positive);

[0068] (2) The model predicts a negative value, but the true value is a positive value, which is a false positive (FP).

[0069] (3) The model predicts a positive value, but the true value is a negative value, which is called a false negative (FN).

[0070] (4) The model predicts a negative value, and the true value is a negative value, which is called True Negative (TN).

[0071] By calculating the confusion matrix of the four types of prediction results, accurate evaluation index values ​​can be obtained. The confusion matrix is ​​as follows: Figure 3 shown.

[0072] The evaluation indicators used in this technology are: Intersection over Union (IoU), Precision (P), Recall (R), and F1-score (F1). Specifically, each evaluation indicator can be expressed by the following formula:

[0073]

[0074] Among them, A represents the pixel area set of the true label, and B represents the pixel area set of the predicted value.

[0075] The model verification process in this embodiment involves an IoU of 75.43%, an accuracy of 85.67%, a recall of 86.81%, and an F1 score of 86.24%. Overall, the construction of the inclusion intelligent recognition model can achieve the expected effect.

[0076] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A fluid inclusion identification method based on convolutional neural network, characterized in that: The steps include: S1, obtaining a target image of a fluid inclusion; S2, establishing a fluid inclusion image data set according to the fluid inclusion target image; S3. Based on the fluid inclusion image dataset, a convolutional neural network model is established to identify and screen the fluid inclusion image dataset.

2. A fluid inclusion identification method based on convolutional neural network according to claim 1, characterized in that: In the step S1, the target image of the fluid inclusion is acquired by using a surface scanning mirror platform, and the spatial position information of the image is recorded simultaneously.

3. The method for fluid inclusion identification based on convolutional neural network according to claim 1, characterized in that: In step S2, the acquired fluid inclusion target image needs to be pixel processed, labeled and data enhanced.

4. The method for fluid inclusion identification based on convolutional neural network according to claim 3, characterized in that: The data enhancement includes operations such as random scaling, random rotation, random horizontal flipping, Gaussian blur, and color gamut transformation on the image.

5. The method for fluid inclusion identification based on convolutional neural network according to claim 1, characterized in that: Before establishing the convolutional neural network model, the image data set needs to be divided into a training set, a validation set, and a test set.

6. The method for fluid inclusion identification based on convolutional neural network according to claim 1, characterized in that: In step S3, the fluid inclusion image data set is processed through an input layer, a convolution layer, a pooling layer, a normalization layer, an activation function, a fully connected layer and an output layer to obtain a convolutional neural network model.

7. The method for fluid inclusion identification based on convolutional neural network according to claim 6, characterized in that: The convolution layer processing method is: C in =N Where: C in The input feature channel dimension of the convolution layer, S is the step size, P is the padding width, H is out is the feature height of the convolutional layer output, W out H is the output feature width of the convolutional layer. in Input feature height for the convolutional layer, W in Input feature width for the convolution layer, N is the number of convolutions.

8. The method for fluid inclusion identification based on convolutional neural network according to claim 1, characterized in that: The method also includes identification verification.

9. The method for fluid inclusion identification based on convolutional neural network according to claim 8, characterized in that: The verification algorithm is as follows: Among them, A represents the pixel area set of the true label, and B represents the pixel area set of the predicted value.

Images