Evaluation method of oil content and maturity of reservoir fluorescence thin sections based on artificial intelligence

Abstract

The present invention provides an artificial intelligence-based method for evaluating the oil content and maturity of reservoir fluorescence thin sections. First, oil-containing fluorescence image data and corresponding RO values ​​are collected to construct an image library. The data is then preprocessed, and features are extracted using a convolutional neural network. Cluster analysis and visualization are performed to identify abnormal images, which are then standardized and saved to minimize errors caused by different environments and equipment. Next, the oil-bearing area is divided, the color space is converted, and extraction is performed based on thresholds. Mask processing and shape filtering are performed to remove oil-immersed minerals. Edges are detected, and the area is extracted and the image is saved. Multidimensional feature fusion extraction is then performed, including the extraction and processing of color, grayscale, and texture features. A 1D convolutional neural network model is then constructed, features are input, and parameters are optimized. Finally, a model is constructed to predict maturity and oil content, and the dimensions are adjusted before input into the model. The present invention combines deep learning with computer vision, offering advantages such as high efficiency, accuracy, and multidimensional fusion.

Description

Technical Field

[0001] The present invention relates to the field of petroleum identification, and in particular to an artificial intelligence-based method for evaluating the oil content and maturity of reservoir fluorescent slices. Background Art

[0002] In traditional manual analysis of crude oil maturity using fluorescence, thin sections of crude oil or asphalt from geological samples are fluorescently stained to emit a specific fluorescence under ultraviolet light, facilitating microscopic observation. Crude oils of varying maturity exhibit different fluorescence colors and intensities, typically shifting from yellow-green to orange-red with increasing maturity. Analysts determine maturity based on fluorescence color, brightness, and texture characteristics, combining this with other information for a comprehensive assessment. Analysts then manually annotate oil-bearing areas and further analyze fluorescence images to complete detailed maturity assessments.

[0003] Existing fluorescence image analysis methods have multiple problems in practical applications. Different analysts may draw different conclusions from the same image, making it difficult to ensure the accuracy and consistency of the results. This difference in human judgment makes the analysis results unstable and increases uncertainty. Manual processing of large numbers of fluorescence images is inefficient. Especially when faced with large-scale oil and gas exploration, traditional methods are cumbersome and time-consuming, and cannot meet the needs of rapid processing. In addition, the entire process is highly dependent on professionals, which not only increases manpower and time costs, but also significantly increases the overall cost of the project. The shortcomings of this existing method are far from the efficient and precise requirements for fine exploration and geochemical research in the information age, which limits its promotion and use in large-scale applications. Summary of the Invention

[0004] The main purpose of the present invention is to provide an artificial intelligence-based method for evaluating the oil content and maturity of reservoir fluorescent thin sections, so as to solve the problems of difficulty in ensuring the accuracy and consistency of analysis results, large differences in human judgment; low efficiency of manual processing, which is difficult to meet the needs of large-scale oil and gas exploration; high reliance on professionals, high manpower and time costs, and a significant increase in the overall cost of the project.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: an artificial intelligence-based reservoir fluorescence thin section oil content and maturity evaluation method, step 1: collecting oil-containing fluorescence image data and the corresponding normalized RO values, and constructing an image library containing different oil source samples;

[0006] Step 2: Preprocess the fluorescence image data, including unifying the image size, converting it to Tensor format, and normalizing it. Use a pretrained convolutional neural network to extract high-level features of the fluorescence image. Apply the KMeans clustering algorithm to analyze the features and perform PCA clustering visualization to identify abnormal images. These abnormal images are then normalized, including brightness normalization, color histogram matching, and contrast enhancement. After image processing is complete, all processed images are saved to a designated directory for use in subsequent model training and testing.

[0007] Step 3: Oil-bearing area segmentation: Convert the RGB image to the HSV color space, extract the oil-bearing area based on the color threshold, perform morphological processing on the extracted oil-bearing area mask, including removing noise and filling holes, perform shape filtering on the mask to remove oil-immersed minerals, perform edge detection on the processed mask, extract the oil-bearing area and mineral area, and save the relevant image;

[0008] Step 4: Multidimensional feature fusion extraction: Extract color features of the oil-bearing area and calculate the mean and standard deviation of each color channel; extract grayscale features, calculate the pixel value mean of the grayscale image, extract texture features, use local binary patterns for feature extraction, and perform histogram statistics and normalization to standardize the features;

[0009] Step 5: Build a 1D convolutional neural network model: Build a 1D convolutional neural network model, input multi-dimensional features, and extract features through multi-layer convolution and pooling operations. Use the Adam optimizer to optimize model parameters to ensure the generalization ability and stability of the model.

[0010] Step 6: Build a model for predicting maturity and oil content from fluorescence images: Adjust the dimension of the feature input to adapt to the model input format, and input the features into the trained 1D convolutional neural network model for maturity prediction.

[0011] In the preferred embodiment, the specific steps in step 2 are:

[0012] A1. According to the width and height of the fluorescence image, the sizes of all images are uniformly processed. In order to ensure the consistency of the input images during training, the entire dataset is first traversed to calculate the maximum width W of all images. max and maximum height H max Each image is padded with black blank areas at the edges to keep its original aspect ratio. Finally, all images will be resized to W. max ×H max Fixed size;

[0013] In this step, the process is as follows:

[0014]

[0015] Where W and H are the width and height of the original image, respectively. Padw and padh are the padding of the image in width and height, respectively. This step ensures that all images fit the network input without changing their proportions.

[0016] A2. Preprocess the image: After padding and adjustment in step 2, the image will be converted to Tensor format and normalized. The normalization formula is as follows:

[0017]

[0018] Where x is the original pixel value, μ is the mean of the image data, and σ is the standard deviation;

[0019] The purpose of normalization is to adjust the image data to preprocess the image;

[0020] After filling and adjusting in step 2, the image will be converted to Tensor format and normalized. The normalization formula is as follows:

[0021]

[0022] Where x is the original pixel value, μ is the mean of the image data, and σ is the standard deviation;

[0023] The purpose of normalization is to adjust the image data to the same scale, usually in the range of [0, 1]. This is achieved by subtracting the mean of each channel of the RGB image and dividing it by the standard deviation, which facilitates subsequent convolutional neural network processing;

[0024] A3. Image feature extraction using the VGG16 model: Features such as color, texture, and shape in fluorescence images are closely related to the maturity and oil content of the sample;

[0025] The core structure of the VGG16 model consists of multiple convolutional layers and pooling layers to extract local and global features of the image; the model processes images through the following structure:

[0026] Conv(W1,H1,D1)→Conv(W2,H2,D2)→…→Dense(4096)→Dense(1000);

[0027] Among them, W n ,H n ,D n Represent the width, height, and depth of the convolution kernel respectively. Each layer of convolution slides on the original image to extract image features at different levels and gradually generate deeper feature representations;

[0028] Before feeding the fluorescence image into the VGG16 model, the image needs to be preprocessed first. The image is resized to the pixel size required by the model and the pixel values ​​are normalized:

[0029]

[0030] Among them, x is the original pixel value, x ′ is the normalized pixel value;

[0031] The convolution operation extracts local features by sliding the convolution kernel on the image. The convolution calculation formula is:

[0032]

[0033] Among them, y i,j,k is the convolution output, x is the input image, w m,n,k is the convolution kernel weight, b k It is bias, which extracts local features of the image by scanning the feature map pixel by pixel;

[0034] After the convolutional layer, the ReLU activation function is applied to introduce nonlinear characteristics, enabling the model to capture complex image features. The formula of ReLU is:

[0035] ReLU(z)=max(0,z);

[0036] The maximum pooling operation is used to further reduce the dimension. The feature map is simplified by selecting the maximum value in the local area of ​​the feature map. The pooling formula is:

[0037]

[0038] in: Represents the value of row i, column j, channel k in the output feature map after the pooling operation; max m,n Indicates the maximum value operation in a local window, the window is usually m×n in size; x i+m-1,j+n-1,k is the value of the input feature map at row i+m-1, column j+n-1, channel k, m and n are the local indices of the sliding window used to select values ​​in the pooling window;

[0039] Finally, after multiple layers of convolution and pooling, VGG16 will expand the generated feature map into a one-dimensional feature vector f(I) with a dimension of 4096, which is used to describe the deep information of the image. f(I) is expressed as:

[0040] f(I)=VGG16(I).

[0041] In the preferred solution, the specific steps of applying the KMeans clustering algorithm to analyze the features in step 2 are:

[0042] B1. The feature vectors extracted from the fluorescence image reflect the important features in the image, and these feature vectors are analyzed using a clustering algorithm;

[0043] Using the KMeans clustering algorithm, the samples are divided into k clusters by minimizing the sum of the squares of the distances from the points in the cluster to the cluster center. The objective function is:

[0044]

[0045] Among them, C i is the i-th cluster, μ i is the cluster center, x j are samples belonging to this cluster;

[0046] Since the feature vector extracted by VGG16 is 4096-dimensional high-dimensional data, in order to understand the clustering results more intuitively, we use principal component analysis, that is, PCA analysis to reduce the high-dimensional features to two or three dimensions. The formula of PCA is: Z = XW;

[0047] Among them, X is the original feature matrix, W is the projection matrix, and Z is the feature matrix after dimensionality reduction. Through PCA, the data can be mapped to a low-dimensional space, and the clustering effect can be displayed through a scatter plot;

[0048] B2. For the standardization of subsequent image processing, select a representative reference image. The selection criteria for the reference image are the images with the Ro value closest to the median.

[0049] First, calculate the median Ro value of all samples using the formula:

[0050]

[0051] Where n is the number of samples, Ro i is the Ro value of the i-th sample;

[0052] By calculating the Ro value, the image with the Ro value closest to the median is selected as the reference image I ref , used for subsequent image standardization processing;

[0053] B3. Different equipment and environmental conditions can affect the performance of fluorescence images. To eliminate these differences, after clustering and comparing with Ro, abnormal ones need to be normalized. By calculating the standard deviation of the Ro value in each cluster, if the standard deviation exceeds the specified threshold, the cluster will be identified as a "differential cluster". Standardization processing includes contrast enhancement, color matching and brightness normalization.

[0054] To improve visual effects, enhance image contrast, make bright areas brighter and dark areas more obvious, and facilitate the extraction of effective features;

[0055] The sharpening filter enhances the edge details in the image, making the texture features of the rock sample more distinct, making it easier for the model to better capture and analyze image features;

[0056] Match the color distribution of all images to be processed with the reference image using the formula:

[0057]

[0058] Among them, I s is the source image, I ref is the reference image, is the histogram matching function;

[0059] Adjust the brightness of the image to make it consistent with the reference image. The brightness normalization formula is:

[0060]

[0061] Among them, I s is the image to be processed, mean(I s ) and mean(I ref ) are the average brightness of the image to be processed and the reference image respectively;

[0062] B4. After image processing is completed, all processed images are saved to the specified directory for use in subsequent model training and testing;

[0063] Based on the clustering results, the processed images are stored in different folders by category. The images in the folders will retain their original file names to ensure that the original data can be easily matched later.

[0064] In order to ensure the processing effect, the comparison results of the original image and the processed image will also be saved, especially the histogram comparison chart, which is used to show the differences and improvements before and after image processing.

[0065] In the preferred embodiment, the specific steps of dividing the oil-bearing area in step 3 are:

[0066] C1. Read and load the result image from the previous step. First, convert the RGB image to the HSV color space. The HSV color space is more suitable for color range screening than RGB. The three channels of the HSV color space represent hue, saturation, and brightness, respectively. In this process, the hue value is obtained by calculating the maximum and minimum values ​​of each pixel and the difference between them. The formula is as follows:

[0067]

[0068] V=max(R,G,B)

[0069] Among them, △ = max(R,G,B)-min(R,G,B). After converting the RGB image to HSV, the color threshold can be used for region screening.

[0070] C2. Define color thresholds to extract oil-containing areas: By setting the HSV ranges of green / cyan and yellow / orange, the color range of the oil-containing area is filtered out. The thresholds are defined as follows:

[0071] HSV range for green / cyan:

[0072] lower_green=[0.25,0.2,0.2], upper_green=[0.5,1.0,1.0]

[0073] HSV range for yellow / orange:

[0074] lower_yellow=[0.12,0.4,0.4], upper_yellow=[0.2,1.0,1.0]

[0075] By screening these two ranges, masks of green / cyan and yellow / orange areas are generated, and then these two masks are merged to obtain a complete oil-bearing area mask;

[0076] C3. Morphological processing of the mask: Morphological operations are used to repair small holes in the mask and remove noise. Morphological operations include binary hole filling, dilation, and erosion.

[0077] The specific steps are as follows:

[0078] Hole filling: Use binary_fill_holes function to fill the holes in the mask, X h =R(B|X); where X h is the filled image, R represents morphological reconstruction;

[0079] Dilation and erosion: Use binary_dilation and binary_erosion functions to dilate and erode the mask to further clean up noise and fill boundaries;

[0080]

[0081] Among them, B is the structural element, Indicates expansion, Indicates corrosion operation;

[0082] The purpose of these morphological operations is to make the extracted oil-bearing areas more complete and clear;

[0083] C4. Shape filtering of the mask to remove oil-immersed minerals: After morphological processing, in order to further remove possible oil-immersed mineral areas, shape features are used for filtering. The specific steps are as follows:

[0084] Circularity calculation: The circularity of each connected area is analyzed to determine whether it is a mineral area. The circularity formula is as follows:

[0085]

[0086] Region division: If the circularity of a region is lower than the specified threshold, it is judged as a mineral region and a mineral region mask is generated; otherwise, the region is retained as an oil-bearing region;

[0087] C5. Perform edge detection on the processed mask: Use the Sobel operator to detect the edge of the oil-bearing area and calculate the horizontal and vertical gradients of each pixel. The formula is as follows:

[0088]

[0089] Among them, G x and G y They are the horizontal and vertical gradients of the image, and the edge image is obtained by calculating their square sum;

[0090] The edge image is further binarized to highlight the boundaries of the oil-bearing area;

[0091] C6. Extracting the oil-containing area: Perform pixel-by-pixel operations on the original image and the processed oil-containing area mask, retaining the original color values ​​of the parts with a value of 1 in the mask, and saving the extracted oil-containing area as a separate image.

[0092] C7. Extracting mineral areas: For areas judged to be minerals, the original color values ​​of the parts with a value of 1 in the mask are also retained, and the extracted mineral areas are saved as separate images.

[0093] In the preferred solution, in steps C1-C7, each time an image is processed, a separate folder is created, and the storage contents include: the original image, the oil-containing area mask, the oil-containing area edge detection result, and the extracted oil-containing area image.

[0094] In the preferred embodiment, the specific method of step 4 is: based on the extracted oil-bearing area image, the color, grayscale and texture features are extracted and the maturity of the image is predicted through the following steps:

[0095] D1. First, convert the image to RGB format and calculate the mean and standard deviation of each color channel; mean ,g mean ,bmean are the average values ​​of red, green and blue channels respectively; r std ,g std ,b std are the standard deviations of the red, green, and blue channels respectively; color feature vector = [r mean ,g mean ,b mean ,r std ,g std ,b std ];

[0096] D2. Convert the image from RGB format to grayscale image and calculate the mean of its pixel values ​​as grayscale feature, grayscale feature = graymean;

[0097] D3. Use local binary pattern (LBP) to extract texture features from grayscale images. The texture features are generated by calculating the histogram of LBP values ​​and normalizing them into probability distribution.

[0098] The formula for LBP calculation is:

[0099]

[0100] Among them, P is the number of sampling points in the neighborhood, g p is the pixel value of each sampling point, g c is the value of the center pixel, and the function s(x) represents the sign function;

[0101] D4. Histogram statistics LBP value and normalization:

[0102]

[0103] D5. Standardization: Use local binary pattern (LBP) to extract texture features from grayscale images. By calculating the histogram of LBP values ​​and normalizing it to a probability distribution, texture features are generated. The standardization formula is as follows:

[0104]

[0105] Among them, x is the original feature and x′ is the standardized feature.

[0106] In the preferred solution, the specific method of the model construction step in step 5 is:

[0107] E1. The input data is the features extracted from the fluorescence image, which mainly includes color features, grayscale features and texture features. The shape of the input feature matrix is ​​(n samples ,n features ,1)where: n samples is the sample size, nfeaturesn is the number of features of each sample, including color, grayscale and texture features;

[0108] E2. The convolution layer is used to extract local features from the input data. Through one-dimensional convolution operations, several feature maps are generated. Each convolution operation performs convolution calculations on the input data and outputs a new feature representation. The formula of the convolution layer is:

[0109]

[0110] Among them, w ij is the convolution kernel weight, x i is the input feature, b j is the bias term, H j is the convolution output feature;

[0111] E3, the pooling layer reduces the dimension of the feature map through downsampling operations while retaining the most significant features. It uses the maximum pooling operation to take the maximum value in the local area. The formula for the pooling operation is:

[0112] P j =max(H j1 ,H j2 ,...,H jn );

[0113] Among them, H ji It is the feature map corresponding to the jth convolution kernel, and the downsampled feature map P is obtained after pooling. j ;

[0114] E4. The fully connected layer flattens the features extracted by the convolution and pooling layers into a vector and linearly combines it with the weights of the fully connected layer to generate the final output. The formula of the fully connected layer is:

[0115]

[0116] Among them, w i is the weight of the fully connected layer, P i is the feature map output by the pooling layer, is the predicted output of the model, that is, the Ro_normalized value of the sample;

[0117] E5. The output layer uses the ReLU activation function to ensure that the predicted value is non-negative and outputs a continuous value to represent the predicted Ro_normalized value:

[0118]

[0119] E6. Use the mean square error as the loss function to measure the difference between the Ro_normalized value predicted by the model and the true value. The formula of the loss function is:

[0120]

[0121] Among them, y i is the true value, is the predicted value, N is the number of samples;

[0122] E7. We use the Adam optimizer, an optimization algorithm with adaptive learning rate, to optimize the model parameters. The parameter update formula of Adam is as follows:

[0123]

[0124] Where: α is the learning rate, and are the bias-corrected first-order and second-order moment estimates of the gradient, and ∈ is a small constant to prevent division by zero.

[0125] In the preferred embodiment, the specific method for constructing the model for predicting maturity and oil content from fluorescence images in step 6 is as follows:

[0126] F1. Input the extracted standardized features into the trained 1D convolutional neural network model to predict the maturity of the fluorescence image;

[0127] Before inputting features into a 1D convolutional neural network, the feature dimensions need to be adjusted to fit the model's input format. For a convolutional neural network, the input data format is usually (number of samples, number of features, 1), so the standardized features need to be expanded in dimension.

[0128] The expanded dimensions are as follows:

[0129] color_features_scaled → (number of samples, 6, 1);

[0130] gray_features_scaled→(number of samples, 1, 1);

[0131] texture_features_scaled→(number of samples, N, 1);

[0132] Where N is the length of the local binary pattern feature, that is, the length of the LBP feature;

[0133] F2. Model prediction: The features processed by the expanded dimension will be passed as input to the trained 1D convolutional neural network model for maturity prediction. The model has been trained to predict the maturity of the fluorescence image based on color, grayscale and texture features. The prediction result of the model will output the predicted maturity value of the image, that is, Ro_normalization. The model prediction formula is:

[0134] Ro归一化 =f(color_features_scaled,gray_features_scaled,texture_features_scaled);

[0135] Among them, f is a 1D convolutional neural network model, Ro 归一化 is the predicted maturity value;

[0136] According to the predicted Ro 归一化 The fluorescence images are divided into the following three maturity levels:

[0137] Low maturity: Ro 归一化 <0.3;

[0138] Medium maturity: 0.3≤Ro 归一化 <0.7;

[0139] High maturity: Ro 归一化 ≥0.7.

[0140] In the preferred embodiment, the method for determining oil content is:

[0141] Oil content assessment: In addition to maturity prediction, the oil content of fluorescent images is assessed using comprehensive indicators such as brightness, texture complexity, and oil-bearing area coverage.

[0142] Oil content judgment index and formula:

[0143] The oil area coverage ratio, namely Oil Area Ratio, is expressed as:

[0144]

[0145] Indicates the proportion of oil-containing area to the entire image;

[0146] Average brightness, or Average Brightness, is expressed as:

[0147]

[0148] Where N represents the total number of pixels in the oil area, and the brightness is extracted from the grayscale image;

[0149] Texture complexity: obtained through local binary pattern texture calculation, that is, LBP texture calculation: Texture complexity = σ (LBP histogram)

[0150] Where σ represents the standard deviation of the LBP histogram, which is used to measure texture complexity.

[0151] In the preferred embodiment, the oil content judgment standard is:

[0152] Based on brightness, texture complexity, and coverage, samples were classified as high, medium, and low oil content;

[0153] High oil content:

[0154] The area ratio of oil-bearing areas is >0.3, which means that the area of ​​oil-bearing areas is large;

[0155] The average brightness of the oil-bearing area is >0.6, indicating strong fluorescence brightness;

[0156] Texture complexity < 0.05, indicating that the fluorescent texture is relatively uniform;

[0157] Judgment logic: Large area, high brightness and uniform texture usually indicate high oil content;

[0158] Medium oiliness:

[0159] The area ratio of oil-bearing areas is >0.1 and ≤0.3, indicating that the area of ​​oil-bearing areas is moderate;

[0160] The average brightness of the oil-bearing area is >0.3 and ≤0.6, indicating moderate fluorescence brightness;

[0161] Texture complexity <0.1, that is, the fluorescent texture is relatively uniform but slightly scattered;

[0162] Judgment logic: Medium area, brightness, and moderate texture complexity usually indicate medium oil content and medium oil and gas generation;

[0163] Low oil content:

[0164] The area ratio of oil-bearing areas is ≤0.1, which means that the area of ​​oil-bearing areas is small;

[0165] The average brightness of the oil-containing area is ≤0.3, indicating low fluorescence brightness;

[0166] Texture complexity ≥ 0.1, indicating that the fluorescence texture is complex and uneven;

[0167] Judgment logic: Small area, low brightness and complex texture usually indicate low oil content.

[0168] The final maturity value and oil content will be output and displayed as the maturity prediction result for the given fluorescence image.

[0169] The present invention provides an artificial intelligence-based method for evaluating the oil content and maturity of reservoir fluorescent thin sections, and a deep learning-based fluorescence image analysis technology: This patent focuses on achieving automated fluorescence image processing and analysis through the combination of deep learning and computer vision.

[0170] Dataset construction and processing method: Label data is generated through standardized image preprocessing and factor analysis to ensure the accuracy and consistency of the dataset.

[0171] Multi-dimensional feature fusion: Comprehensively considers features such as color, brightness, and texture, and performs image analysis from multiple dimensions, enhancing the robustness and reliability of the model.

[0172] Efficiency: By leveraging deep learning for adaptive image enhancement and automating the oil-bearing area identification process, our research significantly improves processing speed, enabling efficient analysis of fluorescence images and significantly improving the efficiency of oil and gas exploration.

[0173] Improved accuracy: The automated process reduces errors caused by human factors. Through refined image processing technology and optimization of deep learning models, it achieves accurate identification of complex oil-bearing areas in fluorescent images, improving the accuracy of crude oil maturity analysis.

[0174] Multi-dimensional feature fusion: Our research uses a multi-dimensional feature fusion approach, combining color, brightness, texture, and other features to construct a comprehensive feature vector that accurately captures the complex properties of crude oil / bitumen in fluorescent thin sections. This fusion strategy significantly improves the accuracy of maturity (Ro) identification compared to single feature analysis.

[0175] Technological innovation and transformation: Our research applies deep learning to image enhancement and automated recognition, enabling adaptive processing and intelligent analysis. This breaks through the limitations of traditional manual analysis and promotes the development of image processing technology and oil and gas exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0176] The present invention will be further described below with reference to the accompanying drawings and examples:

[0177] Figure 1 It is the PCA cluster visualization diagram of the present invention;

[0178] Figure 2 It is a before-and-after comparison of the standardized image of the present invention;

[0179] Figure 3 This is a graph showing changes in the fluorescence image indicators before and after standardization of the present invention;

[0180] Figure 4 It is a schematic diagram of multi-dimensional feature fusion of the present invention;

[0181] Figure 5 It is the feature distribution diagram after PCA dimensionality reduction of the present invention;

[0182] Figure 6 It is the multi-dimensional fusion feature radar chart of the present invention;

[0183] Figure 7 It is a structural diagram of the training model of the present invention;

[0184] Figure 8 is a loss curve diagram of the training model of the present invention;

[0185] Figure 9 This is a flow chart for predicting maturity and oil content according to the present invention;

[0186] Figure 10 It is the present invention that generates a separation region map including edge detection. DETAILED DESCRIPTION

[0187] Example 1

[0188] like Figures 1 to 10 Figure 1 shows an AI-based method for evaluating the oil content and maturity of reservoir fluorescence thin sections. Step 1: Collect oil-bearing fluorescence image data and the corresponding normalized RO values ​​to construct an image library containing samples from different oil source areas. This library should cover a wide range of fluorescence image samples, including samples of varying maturity, oil content, and geological backgrounds, to ensure the broad applicability and accuracy of subsequent analytical models. These fluorescence images will serve as the data foundation for subsequent model training and testing.

[0189] Step 2: Preprocess the fluorescence image data, including unifying the image size, converting it to Tensor format, and normalizing it. Use a pretrained convolutional neural network to extract high-level features of the fluorescence image. Apply the KMeans clustering algorithm to analyze the features and perform PCA clustering visualization to identify abnormal images. These abnormal images are then normalized, including brightness normalization, color histogram matching, and contrast enhancement. After image processing is complete, all processed images are saved to a designated directory for use in subsequent model training and testing.

[0190] Step 3: Oil-bearing area segmentation: Convert the RGB image to the HSV color space, extract the oil-bearing area based on the color threshold, perform morphological processing on the extracted oil-bearing area mask, including removing noise and filling holes, perform shape filtering on the mask to remove oil-immersed minerals, perform edge detection on the processed mask, extract the oil-bearing area and mineral area, and save the relevant image;

[0191] Step 4: Multidimensional feature fusion extraction: Extract color features of the oil-bearing area and calculate the mean and standard deviation of each color channel; extract grayscale features, calculate the pixel value mean of the grayscale image, extract texture features, use local binary patterns for feature extraction, and perform histogram statistics and normalization to standardize the features;

[0192] Step 5: Build a 1D convolutional neural network model: Build a 1D convolutional neural network model, input multi-dimensional features, and extract features through multi-layer convolution and pooling operations. Use the Adam optimizer to optimize model parameters to ensure the generalization ability and stability of the model.

[0193] Step 6: Build a model for predicting maturity and oil content from fluorescence images: Adjust the dimension of the feature input to adapt to the model input format, and input the features into the trained 1D convolutional neural network model for maturity prediction.

[0194] In the preferred embodiment, the specific steps in step 2 are: fluorescence image feature extraction, cluster analysis and image processing, in order to minimize the differences caused by different environments or equipment and prevent the impact on subsequent operations;

[0195] In step 1, a large amount of fluorescence image data has been collected and associated with corresponding Ro values ​​(normalized vitrinite reflectance).

[0196] A1. According to the width and height of the fluorescence image, the sizes of all images are uniformly processed. In order to ensure the consistency of the input images during training, the entire dataset is first traversed to calculate the maximum width W of all images. max and maximum height H max Each image is padded with black blank areas at the edges to keep its original aspect ratio. Finally, all images will be resized to W. max ×H max Fixed size;

[0197] In this step, the process is as follows:

[0198]

[0199] Where W and H are the width and height of the original image, respectively. Padw and padh are the padding of the image in width and height, respectively. This step ensures that all images fit the network input without changing their proportions.

[0200] A2. Preprocess the image: After padding and adjustment in step 2, the image will be converted to Tensor format and normalized. The normalization formula is as follows:

[0201]

[0202] Where x is the original pixel value, μ is the mean of the image data, and σ is the standard deviation;

[0203] The purpose of normalization is to adjust the image data to preprocess the image;

[0204] After filling and adjusting in step 2, the image will be converted to Tensor format and normalized. The normalization formula is as follows:

[0205]

[0206] Where x is the original pixel value, μ is the mean of the image data, and σ is the standard deviation;

[0207] The purpose of normalization is to adjust the image data to the same scale, usually in the range of [0, 1]. This is achieved by subtracting the mean of each channel of the RGB image and dividing it by the standard deviation, which facilitates subsequent convolutional neural network processing;

[0208] A3. Features such as color, texture, and shape in fluorescence images are closely related to the maturity and oil content of the sample. To effectively extract these features, we used the pre-trained convolutional neural network (CNN) VGG16 model. Trained on a large-scale image dataset (ImageNet), the VGG16 model automatically extracts high-level features from fluorescence images. These features better reflect the essential characteristics of the image, such as color and morphology.

[0209] Image feature extraction using the VGG16 model: Features such as color, texture, and shape in fluorescence images are closely related to the maturity and oil content of the sample;

[0210] The core structure of the VGG16 model consists of multiple convolutional layers and pooling layers to extract local and global features of the image; the model processes images through the following structure:

[0211] Conv(W1,H1,D1)→Conv(W2,H2,D2)→…→Dense(4096)→Dense(1000);

[0212] Among them, W n ,H n ,D n Represent the width, height, and depth of the convolution kernel respectively. Each layer of convolution slides on the original image to extract image features at different levels and gradually generate deeper feature representations;

[0213] Before feeding the fluorescence image into the VGG16 model, the image needs to be preprocessed first. The image is resized to the pixel size required by the model and the pixel values ​​are normalized:

[0214]

[0215] Among them, x is the original pixel value, x ′ is the normalized pixel value;

[0216] The convolution operation extracts local features by sliding the convolution kernel on the image. The convolution calculation formula is:

[0217]

[0218] Among them, y i,j,k is the convolution output, x is the input image, w m,n,k is the convolution kernel weight, b k It is bias, which extracts local features of the image by scanning the feature map pixel by pixel;

[0219] After the convolutional layer, the ReLU activation function is applied to introduce nonlinear characteristics, enabling the model to capture complex image features. The formula of ReLU is:

[0220] ReLU(z)=max(0,z);

[0221] The maximum pooling operation is used to further reduce the dimension. The feature map is simplified by selecting the maximum value in the local area of ​​the feature map. The pooling formula is:

[0222]

[0223] in: Represents the value of row i, column j, channel k in the output feature map after the pooling operation; max m,n Indicates the maximum value operation in a local window, the window is usually m×n in size; x i+m-1,j+n-1,k is the value of the input feature map at row i+m-1, column j+n-1, channel k, m and n are the local indices of the sliding window used to select values ​​in the pooling window;

[0224] Finally, after multiple layers of convolution and pooling, VGG16 will expand the generated feature map into a one-dimensional feature vector f(I) with a dimension of 4096, which is used to describe the deep information of the image. f(I) is expressed as:

[0225] f(I)=VGG16(I).

[0226] In the preferred solution, the specific steps of applying the KMeans clustering algorithm to analyze the features in step 2 are:

[0227] B1. The feature vectors extracted from the fluorescence image reflect the important features in the image, and these feature vectors are analyzed using a clustering algorithm;

[0228] Using the KMeans clustering algorithm, the samples are divided into k clusters by minimizing the sum of the squares of the distances from the points in the cluster to the cluster center. The objective function is:

[0229]

[0230] Among them, C i is the i-th cluster, μ i is the cluster center, x j are samples belonging to this cluster;

[0231] Since the feature vector extracted by VGG16 is 4096-dimensional high-dimensional data, in order to understand the clustering results more intuitively, we use principal component analysis, that is, PCA analysis to reduce the high-dimensional features to two or three dimensions. The formula of PCA is: Z = XW;

[0232] Among them, X is the original feature matrix, W is the projection matrix, and Z is the feature matrix after dimensionality reduction. Through PCA, the data can be mapped to a low-dimensional space, and the clustering effect can be displayed through a scatter plot;

[0233] B2. For the standardization of subsequent image processing, select a representative reference image. The selection criteria for the reference image are the images with the Ro value closest to the median.

[0234] First, calculate the median Ro value of all samples using the formula:

[0235]

[0236] Where n is the number of samples, Ro i is the Ro value of the i-th sample;

[0237] By calculating the Ro value, the image with the Ro value closest to the median is selected as the reference image I ref , used for subsequent image standardization processing;

[0238] B3. Different equipment and environmental conditions can affect the performance of fluorescence images. To eliminate these differences, after clustering and comparing with Ro, abnormalities need to be normalized. The standard deviation of the Ro value in each cluster is calculated. If the standard deviation exceeds the specified threshold, the cluster is identified as a "differential cluster". Standardization includes contrast enhancement, color matching, and brightness normalization.

[0239] To improve visual effects, enhance image contrast, make bright areas brighter and dark areas more obvious, and facilitate the extraction of effective features;

[0240] The sharpening filter enhances the edge details in the image, making the texture features of the rock sample more distinct, making it easier for the model to better capture and analyze image features;

[0241] Match the color distribution of all images to be processed with the reference image using the formula:

[0242]

[0243] Among them, I s is the source image, I ref is the reference image, is the histogram matching function;

[0244] Adjust the brightness of the image to make it consistent with the reference image. The brightness normalization formula is:

[0245]

[0246] Among them, I s is the image to be processed, mean(I s ) and mean(I ref ) are the average brightness of the image to be processed and the reference image respectively;

[0247] B4. After image processing is completed, all processed images are saved to the specified directory for use in subsequent model training and testing;

[0248] Based on the clustering results, the processed images are stored in different folders by category. The images in the folders will retain their original file names to ensure that the original data can be easily matched later.

[0249] In order to ensure the processing effect, the comparison results of the original image and the processed image will also be saved, especially the histogram comparison chart, which is used to show the differences and improvements before and after image processing.

[0250] In the preferred embodiment, the specific steps of dividing the oil-bearing area in step 3 are:

[0251] C1. Read and load the result image from the previous step. First, convert the RGB image to the HSV color space. The HSV color space is more suitable for color range filtering than RGB. The three channels of the HSV color space represent hue, saturation, and value. In this process, the hue value is obtained by calculating the maximum and minimum values ​​of each pixel and the difference between them. The formula is as follows:

[0252]

[0253] V=max(R,G,B)

[0254] Among them, △ = max(R,G,B)-min(R,G,B). After converting the RGB image to HSV, the color threshold can be used for region screening.

[0255] C2. Define color thresholds to extract oil-containing areas: By setting the HSV ranges of green / cyan and yellow / orange, the color range of the oil-containing area is filtered out. The thresholds are defined as follows:

[0256] HSV range for green / cyan:

[0257] lower_green=[0.25,0.2,0.2], upper_green=[0.5,1.0,1.0]

[0258] HSV range for yellow / orange:

[0259] lower_yellow=[0.12,0.4,0.4], upper_yellow=[0.2,1.0,1.0]

[0260] By screening these two ranges, masks of green / cyan and yellow / orange areas are generated, and then these two masks are merged to obtain a complete oil-bearing area mask;

[0261] C3. Morphological processing of the mask: Morphological operations are used to repair small holes in the mask and remove noise. Morphological operations include binary hole filling, dilation, and erosion.

[0262] The specific steps are as follows:

[0263] Hole filling: Use binary_fill_holes function to fill the holes in the mask, X h =R(B|X); where X h is the filled image, R represents morphological reconstruction;

[0264] Dilation and erosion: Use binary_dilation and binary_erosion functions to dilate and erode the mask to further clean up noise and fill boundaries;

[0265]

[0266] Among them, B is the structural element, Indicates expansion, Indicates corrosion operation;

[0267] The purpose of these morphological operations is to make the extracted oil-bearing areas more complete and clear;

[0268] C4. Shape filtering of the mask to remove oil-immersed minerals: After morphological processing, in order to further remove possible oil-immersed mineral areas, shape features are used for filtering. The specific steps are as follows:

[0269] Circularity calculation: Analyze the circularity of each connected area to determine whether it is a mineral area. The circularity formula is as follows:

[0270]

[0271] Region division: If the circularity of a region is lower than the specified threshold, it is judged as a mineral region and a mineral region mask is generated; otherwise, the region is retained as an oil-bearing region;

[0272] C5. Perform edge detection on the processed mask: Use the Sobel operator to detect the edge of the oil-bearing area and calculate the horizontal and vertical gradients of each pixel. The formula is as follows:

[0273]

[0274] Among them, G x and G y They are the horizontal and vertical gradients of the image, and the edge image is obtained by calculating their square sum;

[0275] The edge image is further binarized to highlight the boundaries of the oil-bearing area;

[0276] C6. Extracting the oil-containing area: Perform pixel-by-pixel operations on the original image and the processed oil-containing area mask, retaining the original color values ​​of the parts with a value of 1 in the mask, and saving the extracted oil-containing area as a separate image.

[0277] C7. Extracting mineral areas: For areas judged to be minerals, the original color values ​​of the parts with a value of 1 in the mask are also retained, and the extracted mineral areas are saved as separate images.

[0278] In the preferred solution, in steps C1-C7, a separate folder is created each time an image is processed, and the storage contents include: the original image, the oil-containing area mask, the oil-containing area edge detection result, and the extracted oil-containing area image. Figure 10 shown.

[0279] In the preferred solution, the specific method of step four is: multi-dimensional feature fusion extraction (through the extraction of multiple types of different features, it can ensure that the extracted features are accurate, detailed, and have more effective information, preparing for subsequent model training). Based on the extracted oil-bearing area image, the following steps are performed to extract color, grayscale and texture features and predict the maturity of the image.

[0280] Based on the extracted oil-bearing area image, the following steps are performed to extract color, grayscale, and texture features and predict the maturity of the image:

[0281] D1. First, convert the image to RGB format and calculate the mean and standard deviation of each color channel; mean ,g mean ,b mean are the average values ​​of red, green and blue channels respectively; r std ,gstd ,b std are the standard deviations of the red, green, and blue channels respectively; color feature vector = [r mean ,g mean ,b mean ,r std ,g std ,b std ];

[0282] D2. Convert the image from RGB format to grayscale image and calculate the mean of its pixel values ​​as grayscale feature, grayscale feature = graymean;

[0283] D3. Use local binary pattern (LBP) to extract texture features from grayscale images. The texture features are generated by calculating the histogram of LBP values ​​and normalizing them into probability distribution.

[0284] The formula for LBP calculation is:

[0285]

[0286] Among them, P is the number of sampling points in the neighborhood, g p is the pixel value of each sampling point, g c is the value of the center pixel, and the function s(x) represents the sign function;

[0287] D4. Histogram statistics LBP value and normalization:

[0288]

[0289] D5. Standardization: Use local binary pattern (LBP) to extract texture features from grayscale images. By calculating the histogram of LBP values ​​and normalizing it to a probability distribution, texture features are generated. The standardization formula is as follows:

[0290]

[0291] Among them, x is the original feature and x′ is the standardized feature.

[0292] In the preferred embodiment, the specific method for the model construction step in step 5 is to build a 1D convolutional neural network (1D-CNN) model (1D-CNN is considered to have excellent trainability for small-scale datasets and is easy to achieve the purpose with limited sample data). To predict the maturity (Ro_normalization) of the fluorescence image, we use a 1D convolutional neural network (1D-CNN) model. 1D convolutional neural networks are a type of neural network used to process sequence data and are suitable for feature extraction, classification, and regression tasks.

[0293] E1. The input data is the features extracted from the fluorescence image, which mainly include color features (RGB mean and standard deviation), grayscale features (grayscale mean) and texture features (such as local binary pattern LBP). The shape of the input feature matrix is ​​(n samples ,n features ,1)where: n samples is the sample size, n featuresn is the number of features of each sample, including color, grayscale and texture features;

[0294] E2. The convolution layer is used to extract local features from the input data. Through one-dimensional convolution operations, several feature maps are generated. Each convolution operation performs convolution calculations on the input data and outputs a new feature representation. The formula of the convolution layer is:

[0295]

[0296] Among them, w ij is the convolution kernel weight, x i is the input feature, b j is the bias term, H j is the convolution output feature;

[0297] E3, the pooling layer reduces the dimension of the feature map through downsampling operations while retaining the most significant features. It uses the maximum pooling operation to take the maximum value in the local area. The formula for the pooling operation is:

[0298] P j =max(G j1 ,G j2 ,...,G jn );

[0299] Among them, H ji It is the feature map corresponding to the jth convolution kernel, and the downsampled feature map P is obtained after pooling. j ;

[0300] E4. The fully connected layer flattens the features extracted by the convolution and pooling layers into a vector and linearly combines it with the weights of the fully connected layer to generate the final output. The formula of the fully connected layer is:

[0301]

[0302] Among them, w i is the weight of the fully connected layer, P i is the feature map output by the pooling layer, is the predicted output of the model, that is, the Ro_normalized value of the sample;

[0303] E5. The output layer uses the ReLU activation function to ensure that the predicted value is non-negative and outputs a continuous value to represent the predicted Ro_normalized value:

[0304]

[0305] E6. Use the mean square error (MSE) as the loss function to measure the difference between the Ro_normalized value predicted by the model and the true value. The formula of the loss function is:

[0306]

[0307] Among them, y i is the true value, is the predicted value, N is the number of samples;

[0308] E7. We use the Adam optimizer, an optimization algorithm with adaptive learning rate, to optimize the model parameters. The parameter update formula of Adam is as follows:

[0309]

[0310] Where: α is the learning rate, and are the bias-corrected first-order and second-order moment estimates of the gradient, and ∈ is a small constant to prevent division by zero.

[0311] In the preferred embodiment, the specific method for constructing the model for predicting maturity and oil content from fluorescence images in step 6 is as follows: (using the trained model, predict Ro_ 归一化 The extracted standardized features are input into the trained 1D convolutional neural network (CNN) model to predict the maturity of the fluorescence image. The specific steps are as follows:

[0312] F1. Input the extracted standardized features into the trained 1D convolutional neural network model to predict the maturity of the fluorescence image;

[0313] Before inputting features into a 1D convolutional neural network, the feature dimensions need to be adjusted to fit the model's input format. For a convolutional neural network, the input data format is usually (number of samples, number of features, 1), so the standardized features need to be expanded in dimension.

[0314] The expanded dimensions are as follows:

[0315] color_features_scaled → (number of samples, 6, 1);

[0316] gray_features_scaled→(number of samples, 1, 1);

[0317] texture_features_scaled→(number of samples, N, 1);

[0318] Where N is the length of the local binary pattern feature, that is, the length of the LBP feature;

[0319] F2. Model prediction: The features processed by the expanded dimension will be passed as input to the trained 1D convolutional neural network model (CNN) for maturity prediction. The model has been trained to predict the maturity of the fluorescence image based on color, grayscale and texture features. The prediction result of the model will output the predicted maturity value of the image, that is, Ro_normalization. The model prediction formula is:

[0320] Ro 归一化 =f(color_features_scaled,gray_features_scaled,texture_features_scaled);

[0321] Among them, f is a 1D convolutional neural network model (CNN), Ro 归一化 is the predicted maturity value;

[0322] According to the predicted Ro 归一化 The fluorescence images are divided into the following three maturity levels:

[0323] Low maturity: Ro 归一化 <0.3;

[0324] Medium maturity: 0.3≤Ro 归一化 <0.7;

[0325] High maturity: Ro 归一化 ≥0.7.

[0326] In the preferred embodiment, the method for determining oil content is:

[0327] Oil content assessment: In addition to maturity prediction, the oil content of fluorescent images is assessed using comprehensive indicators such as brightness, texture complexity, and oil-bearing area coverage.

[0328] Oil content judgment index and formula:

[0329] The oil-bearing area coverage is expressed as:

[0330]

[0331] Indicates the proportion of oil-containing area to the entire image;

[0332] Average brightness, or Average Brightness, is expressed as:

[0333]

[0334] Where N represents the total number of pixels in the oil area, and the brightness is extracted from the grayscale image;

[0335] Texture complexity: obtained through local binary pattern texture calculation, that is, LBP texture calculation: Texture complexity = σ (LBP histogram)

[0336] Where σ represents the standard deviation of the LBP histogram, which is used to measure texture complexity.

[0337] In the preferred embodiment, the oil content judgment standard is:

[0338] Based on brightness, texture complexity, and coverage, samples were classified as high, medium, and low oil content;

[0339] High oil content:

[0340] The area ratio of oil-bearing areas is >0.3, which means that the area of ​​oil-bearing areas is large;

[0341] The average brightness of the oil-bearing area is >0.6, indicating strong fluorescence brightness;

[0342] Texture complexity < 0.05, indicating that the fluorescent texture is relatively uniform;

[0343] Judgment logic: Large area, high brightness and uniform texture usually indicate high oil content;

[0344] Medium oiliness:

[0345] The area ratio of oil-bearing areas is >0.1 and ≤0.3, indicating that the area of ​​oil-bearing areas is moderate;

[0346] The average brightness of the oil-bearing area is >0.3 and ≤0.6, indicating moderate fluorescence brightness;

[0347] Texture complexity <0.1, that is, the fluorescent texture is relatively uniform but slightly scattered;

[0348] Judgment logic: Medium area, brightness, and moderate texture complexity usually indicate medium oil content and medium oil and gas generation;

[0349] Low oil content:

[0350] The area ratio of oil-bearing areas is ≤0.1, which means that the area of ​​oil-bearing areas is small;

[0351] The average brightness of the oil-containing area is ≤0.3, indicating low fluorescence brightness;

[0352] Texture complexity ≥ 0.1, indicating that the fluorescence texture is complex and uneven;

[0353] Judgment logic: Small area, low brightness and complex texture usually indicate low oil content.

[0354] Output results: The final maturity value and oil content will be output and displayed as the maturity prediction result for the given fluorescence image.

[0355] Example 2

[0356] Further illustrate with reference to Example 1, Figure 1-10 The technical method described in this invention has been implemented in oil and gas exploration in the Ordos Basin, with promising results, for detecting oil content in rock samples. This technology accurately predicts oil content and maturity (Ro) of rock samples through standardized fluorescence image processing, oil-bearing region demarcation, feature extraction, and fusion, combined with a 1D convolutional neural network model. This method effectively addresses the variability of fluorescence images caused by varying equipment and environmental conditions, ensuring consistent analytical results.

[0357] Sample information:

[0358] According to fluorescence analysis reports, the samples came from medium and fine-medium sandstones in wells Yue 73, Yue 79, and Ban 23. Fluorescence manifests as yellow-green or blue-green fluorescence in intergranular pores and on particle surfaces. The intensity of fluorescence is directly correlated with the degree of oil immersion, indicating a high oil content. The technical method of this invention was applied to assess the oil content and maturity of these samples.

[0359] 1. Fluorescent image processing to achieve the same device and environment effect

[0360] Purpose: To eliminate differences in fluorescence images collected under different equipment and environmental conditions, unify image brightness, color, and contrast, and ensure data consistency.

[0361] Operation steps: Use the VGG16 neural network model to extract high-level features of fluorescence images and identify important image features.

[0362] The KMeans clustering algorithm is applied to analyze the features in order to make the high-dimensional features intuitively analyzed, and PCA clustering visualization is used to find abnormal images caused by different equipment and environmental conditions.

[0363] Abnormal images are standardized, including brightness normalization, color histogram matching, and contrast enhancement, so that all images can achieve a uniform presentation effect.

[0364] Table 1. Feature extraction values ​​of VGG16 neural network model

[0365]

[0366] according to Figure 1 , a two-dimensional visualization of the high-dimensional features of fluorescence images after dimensionality reduction using principal component analysis (PCA). The points in the figure are divided into two categories based on the KMeans clustering results, represented by purple and yellow, respectively. Each point in the figure represents an image, and the different colors indicate different classifications of these images in the feature space. The distribution of purple and yellow points is clearly separated, indicating that PCA and cluster analysis can effectively distinguish different categories of fluorescence images. This classification may be related to differences in equipment and environmental conditions, or the oil content and maturity of the samples, indicating that the extracted features can well reflect the differences between images.

[0367] according to Figure 2 、 Figure 3 These metrics reveal changes in image brightness and color distribution after the modification. In particular, the enhancement of the green and blue channels helps to more clearly visualize the fluorescent signal and oil-bearing areas. These adjustments primarily focus on color balance and contrast enhancement, enhancing image visualization and facilitating subsequent oil content analysis. The goal of this processing is to standardize brightness and color, making fluorescent areas more prominent within the image and eliminating interference from varying equipment and environmental conditions, thereby providing consistent input data for subsequent analysis.

[0368] 2. Oil-bearing area division

[0369] Objective: To extract the oil-bearing area from the fluorescence image and demarcate the oil-immersed mineral area through image processing technology to facilitate the subsequent oil content analysis.

[0370] Operation steps: Convert the fluorescence image to HSV color space and set the color threshold according to the specific color (green, cyan, yellow, orange) exhibited by the oil-immersed area in the sample.

[0371] Use morphological operations to remove image noise, fill holes, and ensure that the extracted area is clear and accurate.

[0372] Apply shape feature filtering technology (such as circularity calculation) to distinguish oil-bearing areas from oil-immersed mineral areas, and generate a separation area map including edge detection. The effect diagram is shown as follows Figure 10 shown.

[0373] 3. Multi-dimensional feature fusion

[0374] Objective: To extract and fuse the multidimensional features of oil-bearing areas and generate feature vectors for predicting oil content and maturity (Ro value) of samples.

[0375] Operation steps: Extract three types of features from the extracted oil-bearing area: Color features: RGB mean and standard deviation. Grayscale features: grayscale mean. Texture features: Texture features extracted using Local Binary Patterns (LBP).

[0376] A multi-branch neural network is used to fuse color, grayscale and texture features to generate the final feature vector as the model input. Figure 4 Shown:

[0377] Figure 5 This figure shows the sample distribution after reducing the 128-dimensional features to 2 dimensions. The color of the points indicates the Ro_normalized value: lighter colors indicate higher Ro_normalized values, while darker colors indicate lower values. The figure shows that samples with similar Ro_normalized values ​​tend to cluster, indicating that the reduced features can distinguish different Ro_normalized values ​​to a certain extent. This demonstrates that the information contained in the original high-dimensional features is preserved after dimensionality reduction and that there is a certain correlation with the Ro_normalized value.

[0378] Figure 6 This graph shows the distribution of 128 feature dimensions for three randomly selected samples. Each line represents a sample, with the values ​​of each feature extending radially outward from the center. The graph shows significant differences in the values ​​of the samples across each feature dimension. Some features fluctuate significantly across samples, while others are relatively close, demonstrating both similarities and differences across samples in these features.

[0379] 4.1D Convolutional Neural Network (1D-CNN) Training and Evaluation

[0380] Objective: To train a 1D convolutional neural network model based on fusion features to predict the Ro value of samples and then evaluate their maturity and oil content.

[0381] Operation steps: Build a 1D convolutional neural network model, input multi-dimensional features, and extract features through multi-layer convolution and pooling operations. Use 5-fold cross-validation to train the model to ensure the generalization ability and stability of the model. Evaluate the model on the final test set and calculate its Ro value prediction accuracy. The results are as follows Figure 7-8 exhibit.

[0382] 5. Model Analysis and Results Interpretation

[0383] Objective: To predict the Ro_normalized values ​​of image samples using a trained 1D convolutional neural network model, and interpret the maturity classification and oil content of the samples based on the prediction results.

[0384] Operation steps: Load model and standardization objects: Load the trained model and standardization tools to provide a basis for feature prediction and classification.

[0385] Feature extraction: Extract multidimensional features such as color, grayscale, and texture from the image to form a 128-dimensional feature vector.

[0386] Normalization and Prediction: Normalize the feature vector and use the 1D-CNN model to predict the Ro_normalized value.

[0387] Maturity judgment: Based on the predicted Ro_normalized values, samples were classified into low, medium and high maturity.

[0388] Oil content evaluation: Based on brightness, texture complexity and coverage, it is divided into three categories: high, medium and low. Figure 9 exhibit.

[0389] Table 2 Comparison of predicted and actual maturity and oil content

[0390]

[0391] The final model predicts the Ro value and oil content of samples in the oil-immersed area, and the prediction results are consistent with geological expectations.

[0392] The technical method of the present invention effectively solves the problem of fluorescence image differences under different equipment and environments, and accurately predicts oil content and maturity through multi-dimensional feature fusion and 1D convolutional neural network, providing an important basis for oil source analysis.

[0393] The scope of protection of the present invention shall be the technical solutions described in the claims, including equivalent replacement solutions of the technical features in the technical solutions described in the claims. In other words, equivalent replacement improvements within this scope are also within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based method for evaluating the oil content and maturity of reservoir fluorescent thin sections, characterized by: Step 1: Collect oil-containing fluorescence image data and the corresponding normalized RO values ​​to build an image library containing samples from different oil sources; Step 2: Preprocess the fluorescence image data, including unifying the image size, converting it to Tensor format, and normalizing it. Use a pretrained convolutional neural network to extract high-level features of the fluorescence image. Apply the KMeans clustering algorithm to analyze the features and perform PCA clustering visualization to identify abnormal images. These abnormal images are then normalized, including brightness normalization, color histogram matching, and contrast enhancement. After image processing is complete, all processed images are saved to a designated directory for use in subsequent model training and testing. Step 3: Oil-bearing area segmentation: Convert the RGB image to the HSV color space, extract the oil-bearing area based on the color threshold, perform morphological processing on the extracted oil-bearing area mask, including removing noise and filling holes, perform shape filtering on the mask to remove oil-immersed minerals, perform edge detection on the processed mask, extract the oil-bearing area and mineral area, and save the relevant image; Step 4: Multidimensional feature fusion extraction: Extract color features of the oil-bearing area and calculate the mean and standard deviation of each color channel; extract grayscale features, calculate the pixel value mean of the grayscale image, extract texture features, use local binary patterns for feature extraction, and perform histogram statistics and normalization to standardize the features; Step 5: Build a 1D convolutional neural network model: Build a 1D convolutional neural network model, input multi-dimensional features, and extract features through multi-layer convolution and pooling operations. Use the Adam optimizer to optimize model parameters to ensure the generalization ability and stability of the model. Step 6: Predicting maturity and oil content from fluorescence images: Adjust the feature input dimensions to fit the model input format and input the features into the trained 1D convolutional neural network model for maturity prediction. Oil content judgment The oil content of the fluorescence image was judged by brightness, texture complexity and oil area coverage; The oil content was judged as follows: based on brightness, texture complexity and coverage, the samples were classified as high, medium and low oil content; High oil content: The area proportion of oil-bearing areas is > 0.3; The average brightness of the oil-bearing area is > 0.6; Texture complexity < 0.05; Judgment logic: Larger area, higher brightness and uniform texture indicate higher oil content; Medium oiliness: The area ratio of oil-bearing areas is > 0.1 and ≤ 0.3; The average brightness of the oil-bearing area is > 0.3 and ≤ 0.6; Texture complexity < 0.1; Judgment logic: Medium area, brightness, and moderate texture complexity represent medium oil content and medium oil and gas generation; Low oil content: the area ratio of oil-bearing areas is ≤ 0.1; The average brightness of the oil-bearing area is ≤ 0.3; Texture complexity ≥ 0.1; Judgment logic: small area, low brightness and complex texture indicate low oil content; The final maturity value and oil content will be output and displayed as the maturity prediction result for the given fluorescence image.

2. The artificial intelligence-based reservoir fluorescence thin section oil content and maturity evaluation method according to claim 1 is characterized by: The specific steps in step 2 are: A1. According to the width and height of the fluorescence image, the sizes of all images are uniformly processed. In order to ensure the consistency of the input images during training, the entire dataset is first traversed to calculate the maximum width W of all images. max and maximum height H max Each image is padded with black blank areas at the edges to keep its original aspect ratio. Finally, all images will be resized to W. max ×H max Fixed size; In this step, the process is as follows: Where W and H are the width and height of the original image, respectively. Padw and padh are the padding of the image in width and height, respectively. This step ensures that all images fit the network input without changing their proportions. A2. Preprocess the image: After padding and adjustment in step 2, the image will be converted to Tensor format and normalized. The normalization formula is as follows: Where x is the original pixel value, μ is the mean of the image data, and σ is the standard deviation; The purpose of normalization is to adjust the image data to preprocess the image; After padding and adjustment in step 2, the image will be converted to Tensor format and normalized. The normalization formula is as follows: Where x is the original pixel value, μ is the mean of the image data, and σ is the standard deviation; The purpose of normalization is to adjust the image data to the same scale. The normalization range is [0, 1]. This is achieved by subtracting the mean of each channel of the RGB image and dividing it by the standard deviation, which facilitates subsequent convolutional neural network processing. A3. Image feature extraction using the VGG16 model: The color, texture, and shape features in the fluorescence image are closely related to the maturity and oil content of the sample; The core structure of the VGG16 model consists of multiple convolutional layers and pooling layers to extract local and global features of the image; the model processes images through the following structure: Conv(W1,H1,D1)→Conv(W2,H2,D2)→⋯→Dense(4096)→Dense(1000); Among them, W n ,H n ,D n Represent the width, height, and depth of the convolution kernel respectively. Each layer of convolution slides on the original image to extract image features at different levels and gradually generate deeper feature representations; Before feeding the fluorescence image into the VGG16 model, the image needs to be preprocessed first. The image is resized to the pixel size required by the model and the pixel values ​​are normalized: ； Among them, x is the original pixel value, is the normalized pixel value; The convolution operation extracts local features by sliding the convolution kernel on the image. The convolution calculation formula is: Among them, y i,j,k is the convolution output, x is the input image, w m,n,k is the convolution kernel weight, b k It is bias, which extracts local features of the image by scanning the feature map pixel by pixel; After the convolutional layer, the ReLU activation function is applied to introduce nonlinear characteristics, enabling the model to capture complex image features. The formula of ReLU is: ReLU(z)=max(0,z); The maximum pooling operation is used to further reduce the dimension. The feature map is simplified by selecting the maximum value in the local area of ​​the feature map. The pooling formula is: ； in: Represents the value of row i, column j, channel k in the output feature map after the pooling operation; max m,n Indicates the maximum value operation in the local window, the window size is m×n; x i+m−1,j+n−1,k is the value of the input feature map at row i+m−1, column j+n−1, channel k, m and n are the local indices of the sliding window used to select values ​​in the pooling window; Finally, after multiple layers of convolution and pooling, VGG16 will expand the generated feature map into a one-dimensional feature vector f(I) with a dimension of 4096, which is used to describe the deep information of the image. f(I) is expressed as: f( I )=VGG16( I )。 3. The artificial intelligence-based reservoir fluorescence thin section oil content and maturity evaluation method according to claim 1 is characterized by: The specific steps of applying the KMeans clustering algorithm to analyze the features in step 2 are: B1. The feature vectors extracted from the fluorescence image reflect the important features in the image, and these feature vectors are analyzed using a clustering algorithm; Using the KMeans clustering algorithm, the samples are divided into k clusters by minimizing the sum of the squares of the distances from the points in the cluster to the cluster center. The objective function is: Among them, C i is the i-th cluster, is the cluster center, x j are samples belonging to this cluster; Since the feature vector extracted by VGG16 is 4096-dimensional high-dimensional data, in order to understand the clustering results more intuitively, we use principal component analysis, that is, PCA analysis to reduce the high-dimensional features to two or three dimensions. The formula of PCA is: Z = XW; Among them, X is the original feature matrix, W is the projection matrix, and Z is the feature matrix after dimensionality reduction. Through PCA, the data can be mapped to a low-dimensional space, and the clustering effect can be displayed through a scatter plot; B2. For the standardization of subsequent image processing, select a representative reference image. The selection criteria for the reference image are the images with the Ro value closest to the median. First, calculate the median Ro value of all samples using the formula: Where n is the number of samples, Ro i is the Ro value of the i-th sample; By calculating the Ro value, the image with the Ro value closest to the median is selected as the reference image I ref , used for subsequent image standardization processing; B3. Different equipment and environmental conditions can affect the performance of fluorescence images. To eliminate these differences, after clustering and comparing with Ro, abnormalities require normalization. The standard deviation of the Ro value in each cluster is calculated. If the standard deviation exceeds a specified threshold, the cluster is identified as a "differential cluster." Normalization includes contrast enhancement, color matching, and brightness normalization. To improve visual effects, enhance image contrast, make bright areas brighter and dark areas more obvious, and facilitate the extraction of effective features; The sharpening filter enhances the edge details in the image, making the texture features of the rock sample more distinct, making it easier for the model to better capture and analyze image features; Match the color distribution of all images to be processed with the reference image using the formula: Among them, I s is the source image, I ref is the reference image, is the histogram matching function; Adjust the brightness of the image to make it consistent with the reference image. The brightness normalization formula is: ； Among them, I s is the image to be processed, mean(I s ) and mean(I ref ) are the average brightness of the image to be processed and the reference image respectively; B4. After image processing is completed, all processed images are saved to the specified directory for use in subsequent model training and testing; Based on the clustering results, the processed images are stored in different folders by category. The images in the folders will retain their original file names to ensure that the original data can be matched later. In order to ensure the processing effect, the comparison results of the original image and the processed image will also be saved, and the histogram comparison chart will be used to show the differences and improvements before and after image processing.

4. The artificial intelligence-based reservoir fluorescence thin section oil content and maturity evaluation method according to claim 1 is characterized by: The specific steps for dividing the oil-bearing area in step 3 are: C1. Read and load the result image from the previous step. First, convert the RGB image to the HSV color space. The HSV color space is more suitable for color range screening than RGB. The three channels of the HSV color space represent hue, saturation, and brightness, respectively. In this process, the hue value is obtained by calculating the maximum and minimum values ​​of each pixel and the difference between them. The formula is as follows: in, = max(R, G, B)-min(R, G, B). After converting the RGB image to HSV, the color threshold can be used for region screening. C2. Define color thresholds to extract oil-containing areas: By setting the HSV ranges of green / cyan and yellow / orange, the color range of the oil-containing area is filtered out. The thresholds are defined as follows: HSV range for green / cyan: lower_green = [0.25, 0.2, 0.2], upper_green = [0.5, 1.0, 1.0] HSV range for yellow / orange: lower_yellow = [0.12, 0.4, 0.4], upper_yellow = [0.2, 1.0, 1.0] By screening these two ranges, masks of green / cyan and yellow / orange areas are generated, and then these two masks are merged to obtain a complete oil-bearing area mask; C3. Morphological processing of the mask: Morphological operations are used to repair small holes in the mask and remove noise. Morphological operations include binary hole filling, dilation, and erosion. The specific steps are as follows: Hole filling: Use binary_fill_holes function to fill the holes in the mask, X h =R(B|X); where X h is the filled image, R represents morphological reconstruction; Dilation and erosion: Use binary_dilation and binary_erosion functions to dilate and erode the mask to further clean up noise and fill boundaries; X⊕B={z∣(B) z ∩X=∅}; X⊖B={z∣(B) z ⊆X}; Where B is the structural element, ⊕ represents expansion, and ⊖ represents corrosion operation; The purpose of these morphological operations is to make the extracted oil-bearing areas more complete and clear; C4. Shape filtering of the mask to remove oil-immersed minerals: After morphological processing, in order to further remove possible oil-immersed mineral areas, shape features are used for filtering. The specific steps are as follows: Circularity calculation: The circularity of each connected area is analyzed to determine whether it is a mineral area. The circularity formula is as follows: ； Region division: If the circularity of a region is lower than the specified threshold, it is judged as a mineral region and a mineral region mask is generated; otherwise, the region is retained as an oil-bearing region; C5. Perform edge detection on the processed mask: Use the Sobel operator to detect the edge of the oil-bearing area and calculate the horizontal and vertical gradients of each pixel. The formula is as follows: ； Among them, G x and G y They are the horizontal and vertical gradients of the image, and the edge image is obtained by calculating their square sum; The edge image is further binarized to highlight the boundaries of the oil-bearing area; C6. Extracting the oil-containing area: Perform pixel-by-pixel operations on the original image and the processed oil-containing area mask, retaining the original color values ​​of the parts with a value of 1 in the mask, and saving the extracted oil-containing area as a separate image. C7. Extracting mineral areas: For areas judged to be minerals, the original color values ​​of the parts with a value of 1 in the mask are also retained, and the extracted mineral areas are saved as separate images.

5. The artificial intelligence-based reservoir fluorescence thin section oil content and maturity evaluation method according to claim 4 is characterized by: In steps C1-C7, each time an image is processed, a separate folder is created to store the following contents: the original image, the oil-containing area mask, the oil-containing area edge detection result, and the extracted oil-containing area image.

6. The artificial intelligence-based reservoir fluorescence thin section oil content and maturity evaluation method according to claim 1 is characterized by: The specific method of step 4 is as follows: Based on the extracted oil-bearing area image, the color, grayscale and texture features are extracted and the maturity of the image is predicted through the following steps: D1. First, convert the image to RGB format and calculate the mean and standard deviation of each color channel; mean ,g mean ,b mean are the average values ​​of red, green and blue channels respectively; r std ,g std ,b std are the standard deviations of the red, green, and blue channels respectively; color feature vector = [r mean ,g mean ,b mean ,r std ,g std ,b std ]; D2. Convert the image from RGB format to grayscale image and calculate the mean of its pixel values ​​as grayscale feature, grayscale feature = graymean; D3. Use local binary pattern (LBP) to extract texture features from grayscale images. The texture features are generated by calculating the histogram of LBP values ​​and normalizing them into probability distribution. The formula for LBP calculation is: Among them, P is the number of sampling points in the neighborhood, g p is the pixel value of each sampling point, g c is the value of the center pixel, and the function s(x) represents the sign function; D4. Histogram statistics LBP value and normalization: D5. Standardization: Use local binary pattern (LBP) to extract texture features from grayscale images. By calculating the histogram of LBP values ​​and normalizing it to a probability distribution, texture features are generated. The standardization formula is as follows: Among them, x is the original feature and x′ is the standardized feature.

7. The artificial intelligence-based reservoir fluorescence thin section oil content and maturity evaluation method according to claim 1 is characterized by: The specific method of the model construction step in step 5 is: E1. The input data is the features extracted from the fluorescence image, including color features, grayscale features and texture features. The shape of the input feature matrix is ​​(n samples ,n features ,1)where: n samples is the sample size, n featuresn is the number of features of each sample, including color, grayscale and texture features; E2. The convolution layer is used to extract local features from the input data. Through one-dimensional convolution operations, several feature maps are generated. Each convolution operation performs convolution calculations on the input data and outputs a new feature representation. The formula of the convolution layer is: Among them, w ij is the convolution kernel weight, x i is the input feature, b j is the bias term, H j is the convolution output feature; E3, the pooling layer reduces the dimension of the feature map through downsampling operations while retaining the most significant features. It uses the maximum pooling operation to take the maximum value in the local area. The formula for the pooling operation is: Among them, H ji It is the feature map corresponding to the jth convolution kernel, and the downsampled feature map P is obtained after pooling. j ; E4. The fully connected layer flattens the features extracted by the convolution and pooling layers into a vector and linearly combines it with the weights of the fully connected layer to generate the final output. The formula of the fully connected layer is: Among them, w i is the weight of the fully connected layer, P i is the feature map output by the pooling layer, is the predicted output of the model, that is, the Ro_normalized value of the sample; E5. The output layer uses the ReLU activation function to ensure that the predicted value is non-negative and outputs a continuous value to represent the predicted Ro_normalized value: E6. Use the mean square error as the loss function to measure the difference between the Ro_normalized value predicted by the model and the true value. The formula of the loss function is: Among them, y i is the true value, is the predicted value, N is the number of samples; E7. We use the Adam optimizer, an optimization algorithm with adaptive learning rate, to optimize the model parameters. The parameter update formula of Adam is as follows: Where: α is the learning rate, and are the bias-corrected first-order and second-order moment estimates of the gradient, and ϵ is a small constant to prevent division by zero.

8. The artificial intelligence-based reservoir fluorescence thin section oil content and maturity evaluation method according to claim 1 is characterized by: The specific method for constructing the model for predicting maturity and oil content from fluorescence images in step 6 is as follows: F1. Input the extracted standardized features into the trained 1D convolutional neural network model to predict the maturity of the fluorescence image; Before inputting features into a 1D convolutional neural network, the feature dimensions need to be adjusted to fit the model's input format. For a convolutional neural network, the input data format is (number of samples, number of features, 1), so the standardized features need to be expanded in dimension. The expanded dimensions are as follows: color_features_scaled → (number of samples, 6, 1); gray_features_scaled→(number of samples, 1, 1); texture_features_scaled→(number of samples, N, 1); Where N is the length of the local binary pattern feature, that is, the length of the LBP feature; F2. Model prediction: The features processed by the expanded dimension will be passed as input to the trained 1D convolutional neural network model for maturity prediction. The model has been trained to predict the maturity of the fluorescence image based on color, grayscale and texture features. The prediction result of the model will output the predicted maturity value of the image, that is, Ro_normalization. The model prediction formula is: Ro 归一化 =f(color_features_scaled,gray_features_scaled,texture_features_scaled); Among them, f is a 1D convolutional neural network model, Ro 归一化 is the predicted maturity value; According to the predicted Ro 归一化 The fluorescence images are divided into the following three maturity levels: Low maturity: Ro 归一化 <0.3; Medium maturity: 0.3≤Ro 归一化 <0.7; High maturity: Ro 归一化 ≥0.

7.

9. The artificial intelligence-based reservoir fluorescence thin section oil content and maturity evaluation method according to claim 8, characterized in that: include: Oil content judgment index and formula: The oil-bearing area coverage is expressed as: Indicates the proportion of oil-containing area in the entire image; the average brightness, namely Average Brightness, is expressed as: Where N represents the total number of pixels in the oil area, and the brightness is extracted from the grayscale image; Texture complexity, also known as Texture Complexity: obtained through local binary pattern texture calculation, that is, LBP texture calculation: Texture complexity = σ(LBP histogram), where σ represents the standard deviation of the LBP histogram, which is used to measure texture complexity.

Images