Unconventional reservoir remaining oil intelligent identification and diagnosis method based on ResNet residual network
Through the intelligent recognition method based on ResNet combined with microfluidic oil-driving experimental images, the automatic identification and classification of unconventional reservoir residual oil is solved, and the problem of low accuracy of manual identification in the existing technology is solved, and efficient unconventional reservoir development is achieved .
Patent Information
- Application Number
- CN202411440468.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-16
- Publication Date
- 2024-11-15
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the prior art, the identification and classification of residual oil in unconventional reservoirs mainly relies on manual judgment, which is time-consuming and labor-intensive and has low accuracy, making it difficult to effectively propose targeted procurement measures.
An intelligent identification method based on ResNet residual network is adopted, combined with microfluidic oil-driving experimental images, a residual oil recognition and diagnosis system is established, and the residual oil type is automatically identified through image classification and training models, and corresponding extraction methods are proposed.
The accurate identification and classification of residual oil in unconventional reservoirs is achieved, with an identification accuracy of more than 90%, and the extraction measures for different types of residual oil are provided, which improves the development efficiency of unconventional reservoirs.
Smart Images

Figure CN118968259A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of oil and gas development, and in particular to a method for intelligently identifying and diagnosing residual oil in unconventional reservoirs based on a ResNet residual network. Background Art
[0002] Unconventional oil and gas resources such as tight oil / shale oil are abundant and are an important force in ensuring my country's energy security, but they are difficult to exploit and the microscopic seepage mechanism is unclear. Microfluidics technology, as the main research method for microscopic pore seepage, can simulate the complex fluid behavior in the reservoir and can be used to explore the microscopic multiphase seepage mechanism in the process of unconventional oil and gas resource development and clarify the oil recovery mechanism.
[0003] During the reservoir development process, crude oil in the matrix pores is retained in different forms due to different development methods, forming different types of residual oil. Different types of residual oil require different extraction methods. Therefore, it is necessary to accurately identify the type of residual oil in order to propose corresponding extraction measures. At present, most studies mainly rely on manual subjective judgment to identify the form and distribution of residual oil, which is time-consuming, labor-intensive and has low accuracy. Summary of the invention
[0004] In order to solve the above technical problems, the present invention discloses an intelligent identification and diagnosis method for residual oil in unconventional reservoirs based on ResNet residual network. The method establishes a residual oil identification and diagnosis system based on ResNet and microfluidic oil recovery experimental images, which can accurately identify and classify the residual oil occurrence forms, and propose corresponding recovery methods for the residual oil types, thereby improving the efficient development method of unconventional oil reservoirs.
[0005] To achieve the above object, the present invention adopts the following technical solutions: The intelligent identification and diagnosis method of remaining oil in unconventional reservoirs based on ResNet residual network specifically includes the following steps: Step 1: Use the microscopic visualization experimental platform to obtain experimental images of unconventional reservoirs at different displacement times when different recovery methods are used to obtain the original training image set and the original test image set; Step 2: Use Python to segment the original training image set obtained in step 1, classify and label it after segmentation to obtain the final training image set, and divide the final training image set into a training set and a validation set in a ratio of 4:1; segment the original test image set obtained in step 1 to obtain the final test image set; Step 3: Create a DataLoader to load and process the training set and validation set obtained in step 2, and perform data transformation on them to increase the size of the final training image set to prevent overfitting and increase the generalization of the ResNet101 model. Step 4: Load the Resnet model and define the loss function, optimizer, and learning rate scheduler. Step 5: Execute multiple training cycles, save the training model with the best verification accuracy in each execution cycle, and add early stopping method to prevent overfitting, finally obtain the training model and output the remaining oil classification report; Step 6, loading the final test image set segmented in step 2 into the training model obtained in step 4, automatically identifying and classifying the outputs into different residual oil types, and calculating the oil saturation of different residual oil types; Step seven: Use Python to statistically analyze the utilization degree of different types of residual oil by unconventional reservoir recovery methods, clarify the effects of different recovery methods on the potential of residual oil, and form a diagnostic method for different types of residual oil reservoirs.
[0006] Furthermore, in step 1, the experimental image includes CO 2 Throughput images, water injection throughput images, and well shut-in images.
[0007] Furthermore, in step 2, the image is segmented into 12×24, i.e., 288 small images, and the types of remaining oil are divided into four categories, including: pore throat remaining oil, wall remaining oil, pore throat hydrodynamically retained remaining oil, and microscopic heterogeneous remaining oil.
[0008] Furthermore, in step 3, the image data input into the DataLoader is first resized to 64×64, and then undergoes a series of transformations implemented by the torchvision.transforms module, including random rotation, random cropping and scaling, random saturation adjustment, random horizontal flipping, and conversion to a tensor; the transformed image data is input into the ResNet101 model for normalization.
[0009] Furthermore, in step 4, the Resnet model architecture includes a convolutional layer, a maximum pooling layer, a residual block, a global average pooling layer and a fully connected layer, wherein the convolutional layer uses a 7×7 convolution kernel with a step size of 2 and outputs 64 feature maps; the maximum pooling layer uses a 3×3 pooling kernel with a step size of 2; the residual block contains four residual blocks, each residual block contains multiple residual units, each residual unit contains three convolutional layers, the first residual block contains 3 residual units, the second residual block contains 4 residual units, the third residual block contains 23 residual units, and the fourth residual block contains 3 residual units; the global average pooling layer performs global average pooling on each feature map; the fully connected layer is the output layer, outputting 1000 categories.
[0010] Furthermore, in step 4, the loss function is the cross entropy loss. During the training process, the output of the model and the true label are passed into the cross entropy loss function to calculate the loss. Then, through the back propagation and optimization steps, the parameters of the model are updated to reduce the loss. The control equation is as follows: ; In the formula, p(x i ) is the true distribution, dimensionless; q(x i ) is the predicted distribution, dimensionless; i is the sample number, dimensionless.
[0011] Furthermore, in step 4, the optimizer is a stochastic gradient descent optimizer. The momentum parameter helps the optimizer maintain direction during the optimization process and avoids frequent changes in direction during the gradient descent process, thereby accelerating the convergence speed. The update rules are as follows: ; Where θ is the model parameter, dimensionless; η is the learning rate, dimensionless; J(θ) is the loss function, dimensionless; is the gradient of the loss function with respect to the model parameters θ.
[0012] Furthermore, in step five, each training cycle includes a training phase and a validation phase. In the training phase, the model will traverse all training sets to calculate losses and update parameters. In the validation phase, the validation set is used to evaluate the model performance and the training model with the best validation accuracy is saved. If the validation accuracy improves, the model parameters are saved and training continues. At this time, the performance of the model will gradually improve through multiple training cycles. Otherwise, the counter is increased. If the counter reaches the set patience value, if the validation set accuracy does not improve within 10 training cycles, the training is terminated.
[0013] The beneficial effect of the present invention is that the method focuses on the division and identification of different residual oil formation mechanisms, uses ResNet to perform identification and classification training on the training image set, and the identification accuracy rate reaches more than 90%; This method can automatically identify and classify the remaining oil occurrence forms, diagnose the remaining oil in the reservoir, and propose corresponding potential tapping measures to explore efficient development methods for unconventional oil reservoirs. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 Schematic diagram of a microscopic visualization experimental platform in an embodiment of the present invention; Figure 3 This is a Resnet model architecture diagram in an embodiment of the present invention; Figure 4 The residual oil type identified by the test set in the embodiment of the present invention; In the figure, 1-microinjection pump; 2-microinjector; 3-microfluidic chip; 4-temperature control platform; 5-hot and cold stage; 6-clamp; 7-microscope; 8-CCD camera; 9-computer; 10-valve; 11-light source; 12-back pressure valve; 13-waste liquid bottle. DETAILED DESCRIPTION
[0015] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0016] In the process of reservoir development, the shape of residual oil is currently identified and divided manually. This method requires professional technicians with rich experience, has a certain degree of subjectivity, is time-consuming and labor-intensive, and has low accuracy. The present invention proposes an intelligent identification and diagnosis method for residual oil in unconventional reservoirs based on ResNet residual network to overcome the technical defects of existing microfluidic oil recovery image analysis, to intelligently identify residual oil in unconventional reservoirs, and to propose corresponding recovery measures for specific residual oil types.
[0017] An intelligent identification and diagnosis method for remaining oil in unconventional reservoirs based on ResNet residual network, the specific process is as follows: (1) The microscopic visualization experimental platform is used to obtain experimental images of unconventional reservoir recovery methods at different displacement times, and the original training image set and the original test image set are obtained.
[0018] Experimental images include CO 2 Throughput image, water injection through-put image, well-insulation image. Microscopic visualization experimental platform such as Figure 2 As shown, it includes a micro-injection pump 1, a microfluidic chip 3, a temperature control platform 4, a hot and cold stage 5, a microscope 7, a CCD camera 8, a waste liquid bottle 13, a clamp 6 and a plurality of micro-injectors 2.
[0019] The operation steps of this microscopic visualization experimental platform are as follows: 2 Take throughput as an example, as follows: Step 1: Turn on the light source 11, and inject deionized water into the microfluidic chip 3 through the first microinjector 2 on the microinjection pump 1, and stop injecting deionized water when the microfluidic chip 3 is visually full of water; Step 2: Inject oil into the microfluidic chip 3 through the second microinjector 2 on the microinjection pump 1, and stop injecting oil when the microfluidic chip 3 is visually filled with oil; Step 3: Observe with a microscope 7, select a suitable microscope objective lens, and make the field of view under the microscope eyepiece the microfluidic chip 3; Step 4: Adjust the focal length of the CCD camera 8 so that it focuses on the microfluidic chip 3; Step 5: Control the CCD camera 8 to take an initial photo of the microfluidic chip 3 injected with oil, and transmit it to the data collection system; Step 6: Set the target temperature through the temperature control platform 4, make the hot and cold stage 5 reach the target temperature, and let it stand for ten minutes; Step 7: Close the right valve 18 and inject CO into the microfluidic chip 3 through the third microinjector 2 of the microinjection pump 1 in constant pressure mode. 2 , stabilize for 20 minutes after reaching the set pressure; Step 8: Close the microinjection pump 1, open the left valve 10, open the back pressure valve 12, and inject CO 2 The oil is recovered to the waste liquid bottle 13. During this period, a photo is taken at a frequency of 1 second to collect CO 2 Throughput experiment image.
[0020] (2) Use Python to segment the original training image set, classify and label it after segmentation to obtain the final training image set.
[0021] The image is divided into 12×24, i.e. 288 small images.
[0022] Previous studies have only considered the residual oil form, which resulted in the inability to carry out targeted recovery measures for the residual oil types. This method divides the residual oil types into four categories based on the formation mechanism of the residual oil, including: (a) Pore-throat residual oil refers to the drop-shaped (including columnar) residual oil retained in the pores and throats of the reservoir due to capillary forces; (b) Wall residual oil refers to the continuous and discontinuous film residual oil adsorbed on the mineral surface due to adhesion; (c) Residual oil retained by pore throat hydrodynamics refers to the residual oil in microscopic smooth retention areas such as pore throat corners and blind ends; (d) Micro-heterogeneous remaining oil refers to the remaining oil retained in relatively low-permeability micro-areas due to the micro-heterogeneity of the reservoir.
[0023] (3) Create a DataLoader to load and process the images and labels obtained in step (2), divide them into a training set and a validation set, and perform data transformation to increase the size of the training set to prevent overfitting and increase the generalization of the ResNet101 model.
[0024] Specifically: the input is the image data loaded by the DataLoader. The image data is first resized to 64×64, and then goes through a series of transformations. The transformations are implemented through the torchvision.transforms module, including random rotation, random cropping and scaling, random saturation adjustment, random horizontal flipping, and conversion to tensor; the transformed image data is input into the ResNet101 model for normalization.
[0025] The input is a 4D tensor with a shape of [batch_size, channels, height, width], where batch_size is the number of images in each batch, channels is the number of color channels of the image (for color images, usually 3, corresponding to the three RGB color channels), and height and width are the height and width of the image. In this code, all images are resized to 64×64, so height and width are both 64. Therefore, the input shape is [batch_size, 3, 64, 64].
[0026] (4) Load the Resnet model and define the loss function, optimizer, and learning rate scheduler.
[0027] The Resnet model architecture includes convolutional layers, maximum pooling layers, residual blocks, global average pooling layers and fully connected layers. The convolutional layer uses a 7×7 convolution kernel with a step size of 2 and outputs 64 feature maps. The maximum pooling layer uses a 3×3 pooling kernel with a step size of 2. The residual block contains four residual blocks, each residual block contains multiple residual units, and each residual unit contains three convolutional layers. The first residual block contains 3 residual units, the second residual block contains 4 residual units, the third residual block contains 23 residual units, and the fourth residual block contains 3 residual units. The global average pooling layer performs global average pooling on each feature map. The fully connected layer is the output layer and outputs 1000 categories.
[0028] In each residual unit, the input is first reduced in dimension by a 1×1 convolution layer, then processed by a 3×3 convolution layer, and finally increased in dimension by a 1×1 convolution layer before being added to the original input. To prevent overfitting, the feature value of the last fully connected layer is replaced from 1000 to 512 to improve the generalization ability of the model.
[0029] The loss function in the present invention is cross entropy loss, which is often used in multi-classification problems. During the training process, the output of the model and the true label are passed into the cross entropy loss function to calculate the loss, and then, through back propagation and optimization steps, the parameters of the model are updated to reduce the loss.
[0030] The governing equations are shown below: ; In the formula, p(x i ) is the true distribution, dimensionless; q(x i ) is the predicted distribution, dimensionless; i is the sample number, dimensionless.
[0031] The optimizer is the Stochastic Gradient Descent (SGD), which is one of the most commonly used optimization algorithms. It reduces the amount of calculation by using only one (or a small batch) of samples to calculate the gradient at each iteration. The momentum parameter helps the optimizer maintain direction during the optimization process and avoids frequent changes in direction during the gradient descent process, thereby speeding up convergence. The update rules are as follows: ; Where θ is the model parameter, dimensionless; η is the learning rate, dimensionless; J(θ) is the loss function, dimensionless; is the gradient of the loss function with respect to the model parameters θ.
[0032] Calculate the gradient of the loss function with respect to the current parameters at each step, and then update the parameters one step in the opposite direction of the gradient. The learning rate determines the step size of the update, which is 0.01. Use the Learning Rate Scheduler to multiply the learning rate by 0.1 (gamma=0.1) every 100 steps (step_size=100). This can gradually reduce the learning rate during training, helping the model to optimize more finely when it is close to the optimal solution.
[0033] (5) Execute multiple training cycles, save the training model with the best verification accuracy in each execution cycle, and add early stopping method to prevent overfitting, finally obtain the training model and output the residual oil classification report. Each training cycle includes a training phase and a verification phase. In the training phase, the model will traverse all training sets to calculate the loss and update the parameters. In the verification phase, the verification set is used to evaluate the model performance and save the training model with the best verification accuracy. If the verification accuracy improves, the model parameters are saved and training continues. At this time, the performance of the model will gradually improve through multiple training cycles. Otherwise, the counter is increased. If the counter reaches the set patience value, if the verification set accuracy does not improve within 10 training cycles, the training is terminated.
[0034] (6) The small image segmented in step (2) is loaded into the training model obtained in step (4). The model automatically identifies and classifies the outputs into different residual oil types, and calculates the oil saturation of different residual oil types. Figure 4 The remaining oil types identified in the test set are the final test image set obtained by the segmentation method. The different types of remaining oil are automatically identified and classified and output. Microscopic heterogeneous remaining oil and pore throat remaining oil can be seen in the figure.
[0035] (7) Using Python to calculate CO 2 The degree of utilization of different types of residual oil by unconventional reservoir recovery methods such as huff and puff, water injection huff and puff, and well soaking can be determined. The effects of different recovery methods on the potential of residual oil can be clarified to form a diagnostic method for different types of residual oil reservoirs.
[0036] At present, the shape of residual oil is identified and divided manually, which requires professional and technical personnel with rich experience, has a certain degree of subjectivity, is time-consuming and labor-intensive, and has a low accuracy rate. The present invention divides and identifies different residual oil formation mechanisms, uses ResNet to identify and classify the training image set, and the recognition accuracy rate reaches more than 90%; it can automatically identify and classify the residual oil occurrence form, diagnose the residual oil in the reservoir, and propose corresponding potential tapping measures to explore efficient development methods for unconventional oil reservoirs. The present invention establishes a residual oil identification and diagnosis system based on ResNet and microfluidic oil recovery experimental images, which can accurately identify and classify the residual oil occurrence form, and propose corresponding recovery methods for the residual oil type, to improve the efficient development method of unconventional oil reservoirs.
[0037] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. An intelligent identification and diagnosis method for residual oil in unconventional reservoirs based on ResNet residual network, characterized in that: The specific steps include: Step 1: Use the microscopic visualization experimental platform to obtain experimental images of unconventional reservoirs at different displacement times when different recovery methods are used to obtain the original training image set and the original test image set; Step 2: Use Python to segment the original training image set obtained in step 1, classify and label it after segmentation to obtain the final training image set, and divide the final training image set into a training set and a validation set in a ratio of 4:1; segment the original test image set obtained in step 1 to obtain the final test image set; Step 3: Create a DataLoader to load and process the training set and validation set obtained in step 2, perform data conversion on them to increase the size of the final training image set to prevent overfitting and increase the generalization of the ResNe model. Step 4: Load the Resnet model and define the loss function, optimizer, and learning rate scheduler. Step 5: Execute multiple training cycles, save the training model with the best verification accuracy in each execution cycle, and add early stopping method to prevent overfitting, finally obtain the training model and output the remaining oil classification report; Step 6, loading the final test image set segmented in step 2 into the training model obtained in step 4, automatically identifying and classifying the outputs into different residual oil types, and calculating the oil saturation of different residual oil types; Step seven: Use Python to statistically analyze the utilization degree of different types of residual oil by unconventional reservoir recovery methods, clarify the effects of different recovery methods on the potential of residual oil, and form a diagnostic method for different types of residual oil reservoirs.
2. The method for intelligent identification and diagnosis of remaining oil in unconventional reservoirs based on ResNet residual network according to claim 1, characterized in that: In step 1, the experimental images include CO2 huff and puff images, water injection huff and puff images, and well soaking images.
3. The method for intelligent identification and diagnosis of remaining oil in unconventional reservoirs based on ResNet residual network according to claim 2, characterized in that: In step 2, the image is segmented into 12×24, i.e., 288 small images, and the types of remaining oil are divided into four categories, including: pore throat remaining oil, wall remaining oil, pore throat hydrodynamically retained remaining oil, and microscopic heterogeneous remaining oil.
4. The method for intelligent identification and diagnosis of remaining oil in unconventional reservoirs based on ResNet residual network according to claim 3, characterized in that: In step 3, the image data input into the DataLoader is first resized to 64×64, and then goes through a series of transformations implemented by the torchvision.transforms module, including: random rotation, random cropping and scaling, random saturation adjustment, random horizontal flipping, and conversion to tensor; the transformed image data is input into the ResNet101 model for normalization.
5. The method for intelligent identification and diagnosis of remaining oil in unconventional reservoirs based on ResNet residual network according to claim 4, characterized in that: In step 4, the Resnet model architecture includes a convolutional layer, a maximum pooling layer, a residual block, a global average pooling layer, and a fully connected layer. The convolutional layer uses a 7×7 convolution kernel with a step size of 2 and outputs 64 feature maps. The maximum pooling layer uses a 3×3 pooling kernel with a step size of 2. The residual block contains four residual blocks, each residual block contains multiple residual units, and each residual unit contains three convolutional layers. The first residual block contains 3 residual units, the second residual block contains 4 residual units, the third residual block contains 23 residual units, and the fourth residual block contains 3 residual units. The global average pooling layer performs global average pooling on each feature map. The fully connected layer is the output layer and outputs 1000 categories.
6. The method for intelligent identification and diagnosis of remaining oil in unconventional reservoirs based on ResNet residual network according to claim 5, characterized in that: In step 4, the loss function is the cross entropy loss. During the training process, the model output and the true label are passed into the cross entropy loss function to calculate the loss. Then, through the back propagation and optimization steps, the model parameters are updated to reduce the loss. The control equation is as follows: ; In the formula, p(x i ) is the true distribution, dimensionless; q(x i ) is the predicted distribution, dimensionless; i is the sample number, dimensionless.
7. The method for intelligent identification and diagnosis of remaining oil in unconventional reservoirs based on ResNet residual network according to claim 6, characterized in that: In step 4, the optimizer is a stochastic gradient descent optimizer. The momentum parameter helps the optimizer maintain direction during the optimization process and avoids frequent changes in direction during the gradient descent process, thereby speeding up the convergence speed. The update rules are as follows: ; In the formula, θ is the model parameter, dimensionless; η is the learning rate, dimensionless; J(θ) is the loss function, dimensionless; is the gradient of the loss function with respect to the model parameters θ.
8. The method for intelligent identification and diagnosis of remaining oil in unconventional reservoirs based on ResNet residual network according to claim 7, characterized in that: In step five, each training cycle includes a training phase and a validation phase. In the training phase, the model will traverse all training sets to calculate losses and update parameters. In the validation phase, the validation set is used to evaluate the model performance and the training model with the best validation accuracy is saved. If the validation accuracy improves, the model parameters are saved and training continues. At this time, the performance of the model will gradually improve through multiple training cycles. Otherwise, the counter is increased. If the counter reaches the set patience value, if the validation set accuracy does not improve within 10 training cycles, the training is terminated.
Citation Information
Patent Citations
Residual oil distribution prediction method and device based on deep convolutional neural network
CN113052371A
Training method and system and application method and system of image classification model
CN115601639A
Method and device for determining microcosmic remaining oil quantity in reservoir
CN117333693A
Paddy field weed identification method fusing CNN and Transform dual networks
CN118587596A
