Reservoir oiliness evaluation method and device based on multi-modal data fusion, electronic equipment and storage medium

Abstract

The present application relates to a kind of petroleum exploration technology field, it is a kind of reservoir oiliness evaluation method, device, electronic equipment and storage medium based on multimodal data fusion, including using training sample set to carry out end-to-end joint training to multimodal fusion network, obtain reservoir oiliness evaluation model, wherein multimodal fusion network is double-branch weighted fusion structure, double-branch includes image branch network and mineral branch network;Reservoir oiliness evaluation model is optimized based on verification sample set and test sample set Model parameters.The present application fully considers the potential correlation of cutting fluorescence image and cutting mineral content data, carries out multimodal fusion to cutting fluorescence image and cutting mineral content data, different modal data is processed in parallel by double-branch structure, data features are enriched, feature complementation is realized, and the convergence speed inconsistency problem caused by modal difference is solved by optimizing weighted fusion, improve the accuracy of reservoir oiliness evaluation.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of oil exploration, and is a reservoir oiliness evaluation method and device based on multi-modal data fusion, an electronic device and a storage medium. BACKGROUND

[0002] Reservoir oiliness evaluation is a core link of oil exploration and development, and the main purpose is to provide a scientific basis for subsequent development plan formulation by analyzing the oil and gas occurrence state and distribution characteristics in the reservoir, so as to achieve the purposes of improving exploration efficiency, optimizing development plan, guaranteeing effective utilization of resources and reducing development risks.

[0003] Conventional reservoir oiliness evaluation methods include nuclear magnetic resonance analysis, rock pyrolysis analysis, logging curve analysis and fluorescence content analysis. The nuclear magnetic resonance analysis can quantitatively characterize the oil amount in different occurrence states, and can effectively distinguish movable oil and bound oil by combining T2 cutoff values, but the data is greatly affected by instrument precision and sample preparation; the rock pyrolysis analysis can refine the proportions of free oil and bound oil through segmented pyrolysis, but the drill core samples are easy to lose light hydrocarbons due to exposure; the logging curve analysis can quickly divide oil-bearing layers based on parameters such as resistivity and acoustic time difference, but the oiliness evaluation of fractured reservoirs is distorted due to mud invasion; the fluorescence content analysis is to irradiate the drill core sample with a fluorescent lamp, and the logging geologist observes the color, intensity and distribution area of the fluorescence to qualitatively estimate the oiliness, which relies on the experience of the logging technical personnel, and the cognitive difference of different personnel on the results is large, resulting in a large error in the detection accuracy.

[0004] With the application of machine learning in oilfield intelligentization, machine learning has also been introduced into reservoir oiliness evaluation, for example:

[0005] The existing patent literature one, with the publication number CN118132936A, discloses a multi-classification method for oil reservoirs based on multi-modal fusion, which comprises the following steps: 1) performing data cleaning, abnormal data processing and data imbalance processing operations on the collected conventional logging data; 2) preprocessing and aligning the conventional logging data and the imaging logging data; 3) inputting the conventional logging data features into the XGBoost optimization model branch to perform multi-classification oil reservoir prediction; 4) inputting the imaging logging data into the VIT model branch to perform multi-classification reservoir image recognition; and 5) realizing decision-level fusion of the XGBoost optimization model classification result and the VIT model classification result based on the DS evidence theory. Although the invention patent realizes multi-classification of oil reservoirs based on multi-modal fusion, the network architecture does not consider the modal difference problem, and the feature distribution difference between image data (high-dimensional space data) and conventional logging data (low-dimensional vector) is large, so an adaptive fusion mechanism needs to be designed, and in addition, simple weighted fusion may lead to insufficient interaction between branches, affecting the final accuracy.

[0006] The second existing published patent document, CN103927547B, discloses a method for quickly analyzing rock debris fluorescence images, which includes the following steps: 1) oil component identification training based on an RGB three-dimensional color space to obtain a clustering file; 2) oil component analysis of the rock debris fluorescence image to be tested according to the clustering file obtained in step 1), specifically including the following steps: 2.1) color feature extraction of the input rock debris fluorescence image to be tested, and generation of a feature vector array; 2.2) analysis of each feature vector in the feature vector array; 2.3) until the analysis of all feature vectors in the vector array is completed, the oil components contained in the rock debris fluorescence image and the proportion of each component are calculated. The invention requires manual intervention, and its accuracy and efficiency are limited.

[0007] The third existing published patent document, CN112036430A, discloses a reservoir oiliness prediction method based on a random forest algorithm. Depth domain seismic attribute data is obtained through time-depth conversion; the depth domain seismic attribute data and the oiliness data interpreted from the well log are coarsened into a geological model; the depth domain seismic attribute data is standardized to obtain dimensionless seismic attribute data; the coarsened oiliness data interpreted from the well log and the dimensionless seismic attribute data at the well point are used as training objects, and a training model is formed through a random forest algorithm. The dimensionless seismic attribute data between wells is input into the training model to obtain a three-dimensional reservoir oiliness prediction body; and the best reservoir oiliness prediction body is obtained through optimization. The invention does not involve reservoir oiliness prediction based on multi-modal data. SUMMARY

[0008] The present application provides a reservoir oiliness evaluation method and device based on multi-modal data fusion, electronic equipment and storage medium, which overcomes the shortcomings of the above-mentioned prior art. It effectively solves the problem of the existing reservoir oiliness evaluation method based on fluorescence content, which must manually detect the fluorescence content of each rock debris image, and has high accuracy but low efficiency.

[0009] One of the technical solutions of the present application is achieved by the following measures: a reservoir oiliness evaluation method based on multi-modal data fusion, comprising:

[0010] A training sample set, a verification sample set and a test sample set are established, wherein each sample set is provided with high, medium and low oiliness samples, and each sample includes a rock debris fluorescence image, rock debris mineral content data corresponding to the depth and label information of the oiliness evaluation result;

[0011] The multi-modal fusion network is trained end-to-end jointly by using a training sample set, and the training is ended when a training stop condition is met, and a reservoir oil-bearing property evaluation model is obtained, wherein the multi-modal fusion network is a double-branch weighted fusion structure, and the double-branch includes an image branch network and a mineral branch network;

[0012] The model parameters of the reservoir oil-bearing property evaluation model are optimized based on a verification sample set and a test sample set, and a reservoir oil-bearing property evaluation model meeting the test evaluation requirements is output;

[0013] The reservoir oil-bearing property evaluation model is used for reservoir oil-bearing property evaluation.

[0014] The following is a further optimization or / and improvement of the above technical solutions:

[0015] The multi-modal fusion network includes:

[0016] The image branch network extracts features from the cutting fluorescence image to obtain a corresponding 512-dimensional feature vector, and includes a convolution layer, a maximum pooling layer, four residual blocks and an adaptive layer, and the adaptive layer includes a gradient normalization layer, an adaptive average pooling layer and a full connection layer;

[0017] The mineral branch network extracts features from the cutting mineral content data to obtain a corresponding 512-dimensional feature vector, and includes three full connection layers, and a gradient normalization layer, a BN layer and a Relu activation function are arranged after each full connection layer;

[0018] The feature fusion network fuses the 512-dimensional feature vector output by the image branch network and the 512-dimensional feature vector output by the mineral branch network to obtain a fusion feature;

[0019] The classifier outputs a corresponding reservoir oil-bearing property evaluation result based on the fusion feature.

[0020] The above establishing a training sample set, a verification sample set and a test sample set includes:

[0021] The cutting fluorescence images at multiple depths of a certain historical well are obtained, and the cutting mineral content data at the corresponding depths are determined;

[0022] Each cutting fluorescence image is divided into oil-bearing property based on a fluorescence content threshold to obtain a low oil-bearing property image label data set, a medium oil-bearing property image label data set and a high oil-bearing property image label data set;

[0023] Each cutting fluorescence image is converted into a corresponding gray level histogram, and the gray level histograms in each oil-bearing property image label data set are classified based on histogram correlation;

[0024] Based on the oil-bearing sample data set acquisition rule, low oil-bearing sample data set, medium oil-bearing sample data set, high oil-bearing sample data set are obtained, wherein the oil-bearing sample data set acquisition rule includes that for a certain oil-bearing image label data set, M gray level histograms are selected from each of the corresponding gray level histograms, the detrital fluorescence images corresponding to the selected gray level histograms are retained, the remaining detrital fluorescence images are deleted, and M samples are formed, each sample includes a detrital fluorescence image, a corresponding depth of detrital mineral content data and label information of oil-bearing evaluation results;

[0025] The low oil-bearing sample data set, the medium oil-bearing sample data set and the high oil-bearing sample data set are respectively divided into a training set, a validation set and a test set according to a preset weight, and the sets of the same category are merged to obtain a training sample set, a validation sample set and a test sample set.

[0026] The above also includes preprocessing the training sample set, the validation sample set and the test sample set, wherein the preprocessing includes random cropping, tensor conversion and Z-Score standardization.

[0027] The above uses the reservoir oil-bearing evaluation model to evaluate the reservoir oil-bearing, including:

[0028] The detrital fluorescence image and the corresponding depth of the detrital mineral content data of the reservoir to be evaluated are obtained.

[0029] The detrital fluorescence image and the corresponding depth of the detrital mineral content data are input into the reservoir oil-bearing evaluation model to obtain the reservoir oil-bearing evaluation result.

[0030] The second technical solution of the application is realized by the following measures: a reservoir oil-bearing evaluation device based on multi-modal data fusion, comprising:

[0031] The sample acquisition unit establishes a training sample set, a validation sample set and a test sample set, wherein each sample set is provided with high, medium and low three types of oil-bearing samples, and each sample includes a detrital fluorescence image, a corresponding depth of detrital mineral content data and label information of oil-bearing evaluation results.

[0032] The model training unit uses the training sample set to perform end-to-end joint training on the multi-modal fusion network, and ends the training when the training stop condition is met to obtain the reservoir oil-bearing evaluation model, wherein the multi-modal fusion network is a double-branch weighted fusion structure, and the double-branch includes an image branch network and a mineral branch network.

[0033] The model testing unit optimizes the model parameters of the reservoir oil-bearing evaluation model based on the validation sample set and the test sample set, and outputs the reservoir oil-bearing evaluation model meeting the test evaluation requirements.

[0034] The reservoir oil-bearing property evaluation unit evaluates reservoir oil-bearing property by using a reservoir oil-bearing property evaluation model.

[0035] The following is a further optimization or / and improvement of the above technical solutions:

[0036] The sample acquisition unit comprises:

[0037] The basic data acquisition module acquires the fluorescence images of cuttings at multiple depths of a certain historical well and determines the cuttings mineral content data corresponding to the depths;

[0038] The label data set division module divides each cuttings fluorescence image based on a fluorescence content threshold to obtain a low oil-bearing property image label data set, a medium oil-bearing property image label data set, and a high oil-bearing property image label data set;

[0039] The gray histogram classification module converts each cuttings fluorescence image into a corresponding gray histogram and classifies the gray histograms in each oil-bearing property image label data set based on histogram correlation;

[0040] The sample data set acquisition module obtains a low oil-bearing property sample data set, a medium oil-bearing property sample data set, and a high oil-bearing property sample data set based on oil-bearing property sample data set acquisition rules, wherein the oil-bearing property sample data set acquisition rules comprise, for a certain oil-bearing property image label data set, screening M gray histograms from each type of gray histogram corresponding thereto, retaining the cuttings fluorescence images corresponding to the screened gray histograms, deleting the remaining cuttings fluorescence images, and forming M samples, each of which comprises a cuttings fluorescence image, cuttings mineral content data corresponding to the depth, and label information of an oil-bearing property evaluation result;

[0041] The sample set division module divides the low oil-bearing property sample data set, the medium oil-bearing property sample data set, and the high oil-bearing property sample data set into a training set, a validation set, and a test set according to a preset weight, respectively, merges sets of the same category, and obtains a training sample set, a validation sample set, and a test sample set.

[0042] The reservoir oil-bearing property evaluation unit comprises:

[0043] The to-be-evaluated data acquisition module acquires the fluorescence images of cuttings and the cuttings mineral content data corresponding to the depths of a to-be-evaluated reservoir;

[0044] The oil-bearing property evaluation module inputs the fluorescence images of cuttings and the cuttings mineral content data corresponding to the depths into the reservoir oil-bearing property evaluation model to obtain a reservoir oil-bearing property evaluation result.

[0045] The third technical solution of the present application is realized by the following measures: an electronic device comprising a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to realize the steps in the reservoir oil-bearing property evaluation method based on multi-modal data fusion.

[0046] The fourth technical solution of the present application is realized by the following measures: a storage medium, the storage medium storing a computer program readable by a computer, the computer program being set to execute the steps in the reservoir oil-bearing property evaluation method based on multi-modal data fusion when running.

[0047] The present application does not need to manually detect the fluorescence content of each rock debris image, realizes reservoir oil-bearing property evaluation, improves efficiency on the basis of ensuring high accuracy, and the specific effective effects include:

[0048] The present application fully considers the potential correlation between the rock debris fluorescence image and the rock debris mineral content data, multi-modal fusion of the rock debris fluorescence image and the rock debris mineral content data, parallel processing of different modal data through a double-branch structure, rich data features, complementary features, and solving the inconsistent convergence speed problem caused by modal differences through optimized weighted fusion, so that the reservoir oil-bearing property evaluation model obtained by training can effectively improve the accuracy of reservoir oil-bearing property evaluation.

[0049] The present application uses a gray histogram to select rock debris fluorescence images to construct a training sample set, a verification sample set and a test sample set, so that the coverage of the training sample set, the verification sample set and the test sample set reaches 100%, the high coverage can reduce data bias, the model can more comprehensively learn feature distribution, and the generalization ability and prediction accuracy of the model are improved. BRIEF DESCRIPTION OF DRAWINGS

[0050] APPENDIX Figure 1 An implementation environment schematic diagram is provided for the embodiments of the present application.

[0051] APPENDIX Figure 2 A reservoir oil-bearing property evaluation method flowchart is provided for embodiment 1 of the present application.

[0052] APPENDIX Figure 3 A training sample set, a verification sample set and a test sample set establishment method flowchart is provided for embodiment 2 of the present application.

[0053] APPENDIX Figure 4 A multi-modal fusion network structure schematic diagram is provided for embodiment 3 of the present application.

[0054] APPENDIX Figure 5An example of the gray scale histogram of the fluorescence image of the cuttings in the low, medium and high oil-bearing image label data set provided for the embodiment 5 of the present application is shown in the figure, wherein 5a, 5b and 5c are examples of three types of gray scale histograms in the low oil-bearing image label data set, 5d and 5e are examples of two types of gray scale histograms in the medium oil-bearing image label data set, and 5f, 5g and 5h are examples of three types of gray scale histograms in the low oil-bearing image label data set.

[0055] Figure 6 shows examples of the fluorescence images of the cuttings in the low, medium and high oil-bearing sample data set provided for the embodiment 5 of the present application. Figure 6 Figure 6 shows examples of the fluorescence images of the cuttings in the low, medium and high oil-bearing sample data set provided for the embodiment 5 of the present application.

[0056] Figure 7 shows a comparison of the evaluation results of the reservoir oil-bearing property provided for the embodiment 5 of the present application. Figure 7 Figure 7 shows a comparison of the evaluation results of the reservoir oil-bearing property provided for the embodiment 5 of the present application. Figure 7 Figure 7 shows a comparison of the evaluation results of the reservoir oil-bearing property provided for the embodiment 5 of the present application.

[0057] Figure 8 shows a structure of the device for evaluating the reservoir oil-bearing property provided for the embodiment 6 of the present application. Figure 8 DETAILED DESCRIPTION

[0058] The present application is not limited by the following embodiments, and the specific implementation can be determined according to the technical solutions of the present application and the actual situation.

[0059] Those skilled in the art can understand that, unless specifically stated, the "module" or "unit" in the embodiments of the present application refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory) or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an integral module or unit that includes the functions of the module or unit.

[0060] In addition, "a plurality of" in the embodiments of the present application refers to two or more, and "first" and "second" are used for distinction and cannot be understood as implying relative importance.

[0061] ​The current conventional reservoir oiliness evaluation methods include nuclear magnetic resonance analysis, rock pyrolysis analysis, logging curve analysis, and fluorescence content analysis. The nuclear magnetic resonance analysis can quantitatively characterize the oil amount in different occurrence states, and can effectively distinguish movable oil and bound oil by combining the T2 cutoff value, but the data is greatly affected by the instrument precision and sample preparation; the rock pyrolysis analysis can refine the free oil and bound oil ratio through segmented pyrolysis, but the cutting sample is easy to lose light hydrocarbon due to exposure; the logging curve analysis can quickly divide the oil-bearing layer based on the resistivity and acoustic time difference, but the oiliness evaluation of the fractured reservoir is distorted due to mud invasion; the fluorescence content analysis is to irradiate the cutting sample by a fluorescent lamp, and the logging geologist observes the color, intensity and distribution area of the fluorescence to qualitatively estimate the oiliness, which depends on the experience of the logging technical personnel, and the cognitive difference of different personnel is large, resulting in a large error in the detection accuracy.

[0062] The embodiment of the present application provides a reservoir oiliness evaluation method and device based on multi-modal data fusion, electronic equipment and storage medium, a training sample set, a verification sample set and a test sample set are established, wherein three types of oiliness samples, high, medium and low, are set in each sample set, and each sample includes a cutting fluorescence image, cutting mineral content data corresponding to the depth and label information of the oiliness evaluation result; the multi-modal fusion network is trained end to end by using the training sample set, and the training is ended when the training stop condition is met, and a reservoir oiliness evaluation model is obtained, wherein the multi-modal fusion network is a double-branch weighted fusion structure, and the double-branch includes an image branch network and a mineral branch network; the model parameters of the reservoir oiliness evaluation model are optimized based on the verification sample set and the test sample set, and the reservoir oiliness evaluation model meeting the test evaluation requirements is output.

[0063] The reservoir oiliness evaluation model is used for reservoir oiliness evaluation.

[0064] Among them, the method provided by the embodiment of the present application can involve artificial intelligence (Artificial Intelligence, AI) technology, and can be realized based on artificial intelligence technology, for example, in a deep learning manner, a corresponding model is obtained by using sample training.

[0065] Among them, machine learning (Machine Learning, ML) is a multi-field interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other subjects. It is a subject that studies how computers simulate or realize human learning behavior to obtain new knowledge or skills, and reorganize the existing knowledge structure to continuously improve the performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its application is widespread in various fields of artificial intelligence.

[0066] Deep learning (DL) refers to machine learning based on deep neural network models and methods. It is developed on the basis of statistical machine learning, artificial neural network and other algorithm models, combined with the development of current big data and big computing power. The most important technical feature of deep learning is the ability to automatically extract features.

[0067] The above machine learning and deep learning usually include neural network, belief network, reinforcement learning, transfer learning, inductive learning, and rule teaching learning.

[0068] The loss function is used in the process of training the neural network. Because the output of the neural network is expected to be as close as possible to the value that is actually intended to be predicted, the difference between the predicted value of the current network and the target value is compared, and the weight vector of each layer of the neural network is updated according to the difference between the two. Before the first update, there is usually an initialization process, that is, the parameters of each layer of the neural network are pre-configured. Until the neural network can predict the target value or a value very close to the target value. Therefore, when deep learning is performed, it is necessary to define "how to compare the difference between the predicted value and the target value", which is the loss function.

[0069] As shown in the accompanying Figure 1 , a schematic diagram of an implementation environment provided by an embodiment of the present application is shown. The implementation environment can include a training device and a using device.

[0070] The training device and the using device are computer devices; optionally, the computer device is a terminal device, such as a mobile phone, a tablet computer, a PC (Personal Computer), and the like; or, the computer device is a server, which can be a single server or a server cluster composed of multiple servers, or a cloud computing service center, and the embodiments of the present application do not limit this.

[0071] The training device refers to a computer device with the ability to train and learn a multi-modal fusion network. Optionally, the training device has the ability to obtain a multi-modal fusion network and trains and learns it according to application requirements. For example, the training device obtains a multi-modal fusion network from other devices through a network, and then trains it through training samples according to application requirements, so that the multi-modal fusion network has the ability to obtain a reservoir oiliness evaluation result. Optionally, the training device has the ability to construct a multi-modal fusion network, which can construct a multi-modal fusion network according to application requirements, and then train and learn it. For example, the training device constructs a multi-modal fusion network by itself in order to obtain a reservoir oiliness evaluation result according to a cutting fluorescence image and corresponding cutting mineral content data, and then trains and learns it according to application requirements.

[0072] The device used refers to a computer device that has the requirement to use a multimodal fusion network. Optionally, the device uses the multimodal fusion network from other devices through the network according to the application requirements. For example, if the device has the requirement to evaluate the oil content of a reservoir, it can use the network to obtain a multimodal fusion network for evaluating the oil content of the reservoir after training and learning from other devices, and use the multimodal fusion network to evaluate the oil content of the reservoir.

[0073] Based on this, the technical solution of the present invention will be described and explained below with reference to several examples.

[0074] Example 1: As shown in the attached document Figure 2 As shown in the figure, this invention discloses a method for evaluating reservoir oil-bearing capacity based on multimodal data fusion, comprising:

[0075] Step S110: Establish training sample set, validation sample set, and test sample set. Each sample set contains three types of oil-bearing samples: high, medium, and low. Each sample includes a rock fragment fluorescence image, rock fragment mineral content data at the corresponding depth, and label information of the oil-bearing evaluation result.

[0076] Each of the above samples includes a fluorescent image of rock cuttings, rock cuttings mineral content data at the corresponding depth, and label information for oil-bearing evaluation results. That is, it can be a fluorescent image of rock cuttings at a certain depth of a certain historical well, rock cuttings mineral content data at the corresponding depth, and corresponding oil-bearing evaluation results. The mineral types include quartz, plagioclase, calcite, dolomite, illite, chlorite, montmorillonite, and kaolin.

[0077] Step S120: The multimodal fusion network is jointly trained end-to-end using the training sample set. When the training stopping condition is met, the training ends and the reservoir oil content evaluation model is obtained. The multimodal fusion network is a two-branch weighted fusion structure, and the two branches include an image branch network and a mineral branch network.

[0078] Step S130: Optimize the model parameters of the reservoir oil-bearing evaluation model based on the validation sample set and the test sample set, and output the reservoir oil-bearing evaluation model that meets the test evaluation requirements.

[0079] Step S140: Use the reservoir oil-bearing evaluation model to evaluate the reservoir oil-bearing capacity.

[0080] In practice, the fluorescence content of the cuttings is related to the clay content, cement type and cement content in the reservoir. Specifically, the increase of the clay content will significantly reduce the porosity and permeability of the reservoir, thereby affecting the accumulation and migration of oil and gas, indirectly reducing the fluorescence display intensity and area. There is a more significant correlation between the cement and the fluorescence content of the cuttings in the reservoir. For example, when the content of siliceous (quartz) and argillaceous (chlorite, illite, etc.) cement is high, the pore space will be significantly reduced, and the hydrocarbon accumulation will be inhibited, and the fluorescence display will be weak. However, in addition to the early basement cement, which can greatly reduce the porosity, the late dissolution of calcareous (dolomite, calcite, etc.) cement can locally enhance the oil-bearing property of the reservoir.

[0081] As can be seen from the above, there is a strong correlation between the type and content of cuttings minerals and the fluorescence content of cuttings. Therefore, the embodiments of the present application fully consider the potential correlation between the cuttings fluorescence image and the cuttings mineral content data. While evaluating the oil-bearing property of the reservoir using the cuttings fluorescence image, the cuttings mineral content data is introduced for auxiliary evaluation. Through end-to-end joint training of the double-branch weighted fusion structure of the multi-modal fusion network, an oil-bearing property evaluation model of the reservoir is obtained, which is used for oil-bearing property evaluation of the reservoir, thereby effectively improving the accuracy of the oil-bearing property evaluation of the reservoir without calculating the fluorescence content of the cuttings.

[0082] Further, the above step S140 specifically includes:

[0083] (1) obtaining the cuttings fluorescence image of the reservoir to be evaluated and the cuttings mineral content data corresponding to the depth;

[0084] Here, the cuttings fluorescence image of the reservoir to be evaluated is the cuttings fluorescence image at a certain depth and the cuttings mineral content data corresponding to the depth.

[0085] (2) inputting the cuttings fluorescence image and the cuttings mineral content data corresponding to the depth into the oil-bearing property evaluation model of the reservoir to obtain the oil-bearing property evaluation result of the reservoir.

[0086] Embodiment 2: as shown in the accompanying Figure 3 The embodiments of the present application are further optimized based on the above embodiments, wherein the training sample set, the verification sample set and the test sample set are established, including:

[0087] Step S210, obtaining the cuttings fluorescence image at a plurality of depths of a certain historical well, and determining the cuttings mineral content data corresponding to the depth.

[0088] Step S220, dividing the oil-bearing property of each cuttings fluorescence image based on the fluorescence content threshold to obtain a low oil-bearing property image label data set, a medium oil-bearing property image label data set and a high oil-bearing property image label data set.

[0089] Specifically, the fluorescence content of cuttings at each well depth can be detected, and two fluorescence content thresholds a and b of cuttings are set. The fluorescence images of cuttings at depths corresponding to the fluorescence content of cuttings less than a are classified into the low oil-bearing image label dataset. The fluorescence images of cuttings at depths corresponding to the fluorescence content of cuttings within the interval [a, b] are classified into the medium oil-bearing image label dataset. The fluorescence images of cuttings at depths corresponding to the fluorescence content of cuttings greater than b are classified into the high oil-bearing image label dataset.

[0090] Step S230: Convert each fluorescence image of cuttings into a corresponding grayscale histogram, and classify the grayscale histograms in each oil-bearing image label dataset based on histogram correlation.

[0091] The grayscale histogram has the one-to-many property. That is, due to the position missingness of the grayscale histogram, for multiple different images, as long as their grayscale level frequencies are the same, they all have the same grayscale histogram distribution characteristics. Based on the one-to-many property of the grayscale histogram, it is found that for multiple different images, as long as their grayscale level frequencies are similar, they all have highly similar grayscale histogram distribution characteristics. This "similarity" feature is the theoretical basis for classifying different grayscale histogram types in this embodiment, and the "similarity" can be quantified, that is, compare the grayscale histogram correlation coefficient ( Figure 1 ) with the set threshold to classify the grayscale histogram type. For example, when p > 0.9, it is considered that the similarity of the two grayscale histograms is relatively high and they can be classified into one category. The calculation formula of the correlation coefficient is as follows: p > 0.9, it is considered that the similarity of the two grayscale histograms is relatively high and they can be classified into one category. The calculation formula of the correlation coefficient is as follows:

[0092]

[0093] Among them, p is the grayscale histogram correlation coefficient of two fluorescence images of cuttings, H 1( i ) and H 2( i ) are the statistical values of the two fluorescence images of cuttings in the i th histogram interval respectively; u H1 and u H2 are the means of the two grayscale histograms respectively; k is the number of intervals of the grayscale histogram ( bin quantity).

[0094] Step S230, obtaining low oiliness sample data set, medium oiliness sample data set and high oiliness sample data set based on oiliness sample data set acquisition rule, wherein the oiliness sample data set acquisition rule comprises: for a certain oiliness image label data set, screening M gray level histograms from each of the corresponding gray level histograms, retaining the lithic fluorescence images corresponding to the screened gray level histograms, deleting the remaining lithic fluorescence images, and forming M samples, each sample comprising a lithic fluorescence image, lithic mineral content data corresponding to the depth and label information of the oiliness evaluation result;

[0095] For example, there are three types of gray level histograms A, B and C in the low oiliness sample data set, M gray level histograms are selected from each type of gray level histogram, and the lithic fluorescence images corresponding to each gray level histogram are obtained, and the lithic mineral content data corresponding to the depth of the lithic fluorescence images are obtained, thereby constituting M samples, each sample comprising a lithic fluorescence image, lithic mineral content data corresponding to the depth and label information of the oiliness evaluation result.

[0096] Step S240, dividing the low oiliness sample data set, the medium oiliness sample data set and the high oiliness sample data set into a training set, a validation set and a test set according to a preset weight, and merging the sets of the same category to obtain a training sample set, a validation sample set and a test sample set.

[0097] For example, if the weight is set to 0.7, 0.1 and 0.2, 70% of the samples in the low oiliness sample data set are used as the training set, 10% of the samples are used as the validation set, and 20% of the samples are used as the test set, and the same applies to the medium oiliness sample data set and the high oiliness sample data set; after being divided into the training set, the validation set and the test set, the training sets in the low oiliness sample data set, the medium oiliness sample data set and the high oiliness sample data set are merged to form the training sample set, and the same applies to the merging of the validation sample set and the test sample set.

[0098] Step S250, preprocessing the training sample set, the validation sample set and the test sample set, wherein the preprocessing comprises random cropping, tensor conversion and Z-Score standardization.

[0099] Specifically, the lithic fluorescence image is adjusted to 256x256 pixels in size, then a 224x224 image is cropped from the center of the lithic fluorescence image, the cropped lithic fluorescence image is converted into a tensor, and finally the processed training sample set, the validation sample set and the test sample set are subjected to Z-Score standardization.

[0100] Embodiment 3: as shown in the accompanying Figure 4 The embodiment of the present application is a further optimization of the above-mentioned embodiment, wherein the multi-modal fusion network comprises:

[0101] The image branch network extracts features of the cutting fluorescence image to obtain a corresponding 512-dimensional feature vector, and specifically includes a convolution layer, a maximum pooling layer, four residual blocks, and an adaptive layer. The adaptive layer includes a gradient normalization layer, an adaptive average pooling layer, and a full connection layer.

[0102] Specifically, the image branch network of the embodiment is a ResNet34 residual network image processing channel. The main body adopts a ResNet34 network backbone, retains the first four residual blocks, removes the original global average pooling layer and full connection layer, and adds an adaptive layer. The adaptive layer specifically includes: the convolution layer can be a 7x7 convolution layer for initially extracting an image feature map; the maximum pooling layer selects the maximum value in a local region of the input image feature map as the output, realizes feature screening and dimension compression, and enhances the translation invariance and computational efficiency of the model; the four residual blocks each include at least two 3x3 convolution layers, one of which is followed by a batch normalization (Batch Normalization) and a ReLU activation function; the adaptive layer includes a gradient normalization layer, an adaptive average pooling layer, a full connection layer, a batch normalization layer (Batch Normalization, BN), and a ReLU activation function. Specifically, the gradient normalization layer balances the gradient scale during backpropagation, ensures the consistency of the gradient distribution of different samples, alleviates the information loss bias after pooling, and avoids the dominant role of deep convolution features in updating the subsequent full connection layer. The specific mathematical formula is as follows: the adaptive average pooling layer compresses the feature map size to 1x1 and outputs a 512-dimensional feature vector; the full connection layer connects each neuron of a layer to all neurons of the next layer to realize information transmission through a weight matrix; the batch normalization layer optimizes the training stability; and the ReLU activation function enhances the expression ability of the model.

[0103]

[0104] wherein g is an original gradient tensor; ||g||2 is an L2 norm of the gradient tensor; ε is a very small constant, which can be 10 to 6; and g' is a normalized gradient tensor.

[0105] The mineral branch network extracts features of the cutting mineral content data to obtain a corresponding 512-dimensional feature vector, and specifically includes three full connection layers, each of which is followed by a gradient normalization layer, a BN layer, and a ReLU activation function.

[0106] Specifically, the mineral branch network of the embodiment is an MLP (Multi-Layer Perceptron) mineral data processing channel, wherein three fully connected layers are respectively a fully connected layer (FC1), a fully connected layer (FC2), and a fully connected layer (FC3). The fully connected layer (FC1) specifically expands the dimension of the mineral content numerical value to 128 dimensions, the fully connected layer (FC2) expands the output dimension to 256 dimensions, and the fully connected layer (FC3) expands the output dimension to 512 dimensions and outputs. A gradient normalization layer, a BN layer, and a Relu activation function are arranged after each fully connected layer. Each neuron of the fully connected layer is connected to all neurons of the next layer, and information transmission is realized through a weight matrix. The batch normalization layer optimizes the training stability and accelerates the training convergence. The Relu activation function enhances the model expression capability.

[0107] (Three) Feature fusion network, fuse the 512-dimensional feature vector output by the image branch network and the 512-dimensional feature vector output by the mineral branch network to obtain a fused feature.

[0108] Specifically, the feature fusion network is constructed using a self-attention mechanism to realize dynamic weight allocation. The specific fusion process includes:

[0109] (1) Feature alignment and projection: L2 normalization is performed on the two output 512-dimensional feature vectors output by the image branch network and the mineral branch network to eliminate the feature scale difference between the modalities in the forward propagation, and then projection operation is performed to map the two modal features to shared query (Q), key (K), and value (V) spaces: Query , Q 、 Key , K 、 Value , V 、

[0110] Q = W Q F img , K = W K F mineral , V = W V F mineral

[0111] , W Q 、 W K 、 W V is a weight matrix learned through training; F img is a feature output by the image branch network; Fmineral is the feature output by the mineral branch network.

[0112] (2) Calculate the attention weight, calculate the correlation between the feature vector output by the image branch network and the feature vector output by the mineral branch network through dot product (Attention Score, AS), and then use the Softmax function to generate the attention weight matrix (A):

[0113]

[0114] A = Softmax AS

[0115] wherein, K T is the transpose of the key matrix; d k = 512, which is the feature dimension scaling factor.

[0116] (3) Weighted average fusion, the two modal feature vectors (the feature vector output by the image branch network and the feature vector output by the mineral branch network) are weighted and averaged to form a fusion feature vector using the attention weight, and the weight matrix A is automatically updated through back propagation, and the initial weight is 0.5;

[0117]

[0118] wherein, F fusion is the fusion feature vector.

[0119] (4) Classifier, based on the fusion feature, output the corresponding reservoir oiliness evaluation result.

[0120] Specifically, the input of the classifier in the embodiment is a 512-dimensional fusion feature vector, and the output is a full connection layer (512, C), which maps the fusion feature to the final class number C, and uses the Softmax activation function to obtain the output class probability (i.e. the reservoir oiliness evaluation result).

[0121] Embodiment 4: The embodiment of the present application is a further optimization of the above-mentioned embodiment, wherein the construction process of the reservoir oiliness evaluation model comprises:

[0122] Step S310, establish a training sample set, a verification sample set and a test sample set, wherein each sample set is provided with high, medium and low three types of oiliness samples, and each sample includes a rock debris fluorescence image, a corresponding depth of rock debris mineral content data and label information of oiliness evaluation result;

[0123] ​​Each of the above samples includes a cutting fluorescence image, cutting mineral content data corresponding to the depth, and label information of the oil-bearing property evaluation result, that is, the cutting fluorescence image, the cutting mineral content data, and the corresponding oil-bearing property evaluation result of a certain historical well at a certain well depth.

[0124] In step S320, the multi-modal fusion network is trained end-to-end jointly using the training sample set, and when a training stop condition is met, the training is ended, and a reservoir oil-bearing property evaluation model is obtained, wherein the multi-modal fusion network is a double-branch weighted fusion structure, and the double-branch includes an image branch network and a mineral branch network.

[0125] The training stop condition includes a loss function and a maximum number of iterations. If the loss function requirement is met, the training is stopped. If the loss function requirement is not met but the maximum number of iterations is met, the training is stopped.

[0126] The parameter setting of the model during the training can include: all fully connected layers are initialized by He to obtain faster convergence; the optimizer is selected as Adam, the initial learning rate of the image branch network is set as 1e-5, and the initial learning rate of the mineral branch network is set as 1e-4; the loss function is Cross-Entropy Loss; the learning rate scheduling adopts a cosine annealing strategy, the cosine period is 30 epochs, and the minimum learning rate is 1e-6; the number of training rounds is 100 epochs, and the Batch Size is 32.

[0127] It should be further noted that the training process of the embodiment adopts an end-to-end joint training strategy to realize synchronous optimization of multi-module parameters. The image branch network and the mineral branch network simultaneously receive input data of their respective modalities. The image branch network can be pre-trained using an ImageNet data set before training, all residual blocks are frozen, and only the adaptation layer, the mineral branch network, the feature fusion network, and the classifier are trained in subsequent training.

[0128] In step S330, the model parameters of the reservoir oil-bearing property evaluation model are optimized based on the verification sample set and the test sample set, and a reservoir oil-bearing property evaluation model meeting the test evaluation requirements is output.

[0129] The early stopping mechanism is introduced during the verification. If the accuracy of the verification sample set does not improve for 15 consecutive epochs, the training is terminated, and the model is saved.

[0130] The confusion matrix can be introduced during the test to evaluate the effectiveness of the reservoir oil-bearing property evaluation model using the accuracy.

[0131] Embodiment 5: The effectiveness of the reservoir oil-bearing property evaluation method based on multi-modal data fusion disclosed in the present application is verified by using the embodiment of the present application, and the specific process is as follows:

[0132] (1) Obtain multiple fluorescence images of cuttings and cuttings mineral content data at corresponding depths at the depth of 4216.0m to 5216.0m of X well, and the mineral types include quartz, plagioclase, calcite, dolomite, illite, chlorite, montmorillonite, and kaolin.

[0133] For example, at the depth of 4216.0m to 5216.0m, one fluorescence image of cuttings is taken every 2 meters, and the cuttings mineral content at the corresponding depth is tested.

[0134] (2) Set the cuttings fluorescence content thresholds a and b as 8% and 27% respectively, set 20 gray scale histograms in each type of gray scale histogram, set the weights for dividing the training set, the validation set, and the test set as 0.7, 0.1, and 0.2, and then establish the training sample set, the validation sample set, and the test sample set according to the method steps of Example 2.

[0135] Among them, based on the oil-bearing sample data set acquisition rule, the low oil-bearing sample data set, the medium oil-bearing sample data set, and the high oil-bearing sample data set can be obtained as follows:

[0136] The cuttings fluorescence images at the depths corresponding to the cuttings fluorescence content of <8% are divided into the low oil-bearing image label data set, and the gray scale histograms in the low oil-bearing image label data set are divided into three categories based on histogram correlation (exhibits a, 5b, and 5c show examples of three types of gray scale histograms) (if two gray scale histogram correlation coefficients are >0.9, they are divided into one category), 20 gray scale histograms are selected from each type of gray scale histogram, the cuttings fluorescence images corresponding to the selected gray scale histograms are retained, and the remaining cuttings fluorescence images are deleted to form 60 samples (exhibits a, 6b, and 6c show examples of the retained 3 cuttings fluorescence images), and the low oil-bearing sample data set is formed. Figure 5 Figure 6 The cuttings fluorescence images at the depths corresponding to the cuttings fluorescence content of <8% are divided into the low oil-bearing image label data set, and the gray scale histograms in the low oil-bearing image label data set are divided into three categories based on histogram correlation (exhibits a, 5b, and 5c show examples of three types of gray scale histograms) (if two gray scale histogram correlation coefficients are >0.9, they are divided into one category), 20 gray scale histograms are selected from each type of gray scale histogram, the cuttings fluorescence images corresponding to the selected gray scale histograms are retained, and the remaining cuttings fluorescence images are deleted to form 60 samples (exhibits a, 6b, and 6c show examples of the retained 3 cuttings fluorescence images), and the low oil-bearing sample data set is formed.

[0137] The cuttings fluorescence images at the depths corresponding to the cuttings fluorescence content of <8% are divided into the low oil-bearing image label data set, and the gray scale histograms in the low oil-bearing image label data set are divided into three categories based on histogram correlation (exhibits a, 5b, and 5c show examples of three types of gray scale histograms) (if two gray scale histogram correlation coefficients are >0.9, they are divided into one category), 20 gray scale histograms are selected from each type of gray scale histogram, the cuttings fluorescence images corresponding to the selected gray scale histograms are retained, and the remaining cuttings fluorescence images are deleted to form 60 samples (exhibits a, 6b, and 6c show examples of the retained 3 cuttings fluorescence images), and the low oil-bearing sample data set is formed. Figure 5 Figure 6 The cuttings fluorescence images at the depths corresponding to the cuttings fluorescence content of <8% are divided into the low oil-bearing image label data set, and the gray scale histograms in the low oil-bearing image label data set are divided into three categories based on histogram correlation (exhibits a, 5b, and 5c show examples of three types of gray scale histograms) (if two gray scale histogram correlation coefficients are >0.9, they are divided into one category), 20 gray scale histograms are selected from each type of gray scale histogram, the cuttings fluorescence images corresponding to the selected gray scale histograms are retained, and the remaining cuttings fluorescence images are deleted to form 60 samples (exhibits a, 6b, and 6c show examples of the retained 3 cuttings fluorescence images), and the low oil-bearing sample data set is formed.

[0138] ​​Rock cutting fluorescence images at depths corresponding to >27% rock cutting fluorescence content were assigned to the high oil-bearing image label dataset. Based on histogram correlation, the gray-level histograms in the high oil-bearing image label dataset were divided into three categories (see attached). Figure 5 f, 5g, and 5h show examples of three types of grayscale histograms. From each type of grayscale histogram, 20 grayscale histograms are selected. The corresponding rock fragment fluorescence images for each selected grayscale histogram are retained, and the remaining rock fragment fluorescence images are deleted, resulting in 60 samples (see attached). Figure 6 g, 6h, and 6i show three examples of retained rock debris fluorescence images, forming a high-oil-content sample dataset.

[0139] (3) The parameter settings of the model include: all fully connected layers are initialized with He to achieve faster convergence; the optimizer is Adam, the initial learning rate of the image branch network is set to 1e-5, and the initial learning rate of the mineral branch network is 1e-4; the loss function is Cross-Entropy Loss; the learning rate scheduling adopts the cosine annealing strategy, the cosine period is 30 epochs, and the minimum learning rate is 1e-6; the number of training epochs is 100 epochs, and the batch size is 32.

[0140] (4) The model is trained using the model training method described in the above embodiment. The rock cutting images and mineral content data of the test sample set are input into the reservoir oil-bearing evaluation model to obtain the corresponding reservoir oil-bearing evaluation results, which are shown in the attached figure. Figure 7 The confusion matrix shown in a, appendix Figure 7 b shows the results of using the commonly used image segmentation method to calculate fluorescence content and then evaluate reservoir oil content. Comparison with the test sample set results shows that the reservoir oil content evaluation method based on multimodal data fusion disclosed in this invention can improve the accuracy by 9.38%, which significantly improves the evaluation accuracy.

[0141] Example 6: As shown in the appendix Figure 8 As shown in the figure, an embodiment of the present invention discloses a reservoir oil-bearing evaluation device based on multimodal data fusion, comprising:

[0142] The sample acquisition unit establishes a training sample set, a validation sample set, and a test sample set. Each sample set contains three types of oil-bearing samples: high, medium, and low. Each sample includes a rock fragment fluorescence image, rock fragment mineral content data at the corresponding depth, and label information for the oil-bearing evaluation result.

[0143] The model training unit uses the training sample set to perform end-to-end joint training of the multimodal fusion network. When the training stopping condition is met, the training ends and the reservoir oil content evaluation model is obtained. The multimodal fusion network is a two-branch weighted fusion structure, and the two branches include an image branch network and a mineral branch network.

[0144] The model testing unit optimizes the model parameters of the reservoir oil-bearing property evaluation model based on the verification sample set and the test sample set, and outputs a reservoir oil-bearing property evaluation model meeting the test evaluation requirements.

[0145] The reservoir oil-bearing property evaluation unit performs reservoir oil-bearing property evaluation using the reservoir oil-bearing property evaluation model.

[0146] The sample acquisition unit includes:

[0147] The basic data acquisition module acquires the fluorescence images of cuttings at multiple depths of a certain historical well and determines the cuttings mineral content data corresponding to the depths.

[0148] The label data set division module divides each cuttings fluorescence image based on the fluorescence content threshold to obtain a low oil-bearing property image label data set, a medium oil-bearing property image label data set, and a high oil-bearing property image label data set.

[0149] The gray histogram classification module converts each cuttings fluorescence image into a corresponding gray histogram and classifies the gray histograms in each oil-bearing property image label data set based on histogram correlation.

[0150] The sample data set acquisition module obtains a low oil-bearing property sample data set, a medium oil-bearing property sample data set, and a high oil-bearing property sample data set based on oil-bearing property sample data set acquisition rules, wherein the oil-bearing property sample data set acquisition rules include, for a certain oil-bearing property image label data set, screening M gray histograms from each type of gray histogram corresponding thereto, retaining the cuttings fluorescence images corresponding to the screened gray histograms, deleting the remaining cuttings fluorescence images, and forming M samples, each of which includes a cuttings fluorescence image, cuttings mineral content data corresponding to the depth, and label information of the oil-bearing property evaluation result.

[0151] The sample set division module divides the low oil-bearing property sample data set, the medium oil-bearing property sample data set, and the high oil-bearing property sample data set into a training set, a verification set, and a test set according to a preset weight, respectively, merges sets of the same category, and obtains a training sample set, a verification sample set, and a test sample set.

[0152] The reservoir oil-bearing property evaluation unit includes:

[0153] The data to be evaluated acquisition module acquires the cuttings fluorescence images of the reservoir to be evaluated and the cuttings mineral content data corresponding to the depth.

[0154] The oil-bearing property evaluation module inputs the cuttings fluorescence images and the cuttings mineral content data corresponding to the depth into the reservoir oil-bearing property evaluation model to obtain the reservoir oil-bearing property evaluation result.

[0155] In the embodiment 7, a storage medium is disclosed. The storage medium stores a computer program readable by a computer. The computer program is configured to execute a reservoir oil-bearing property evaluation method based on multi-modal data fusion when running.

[0156] The storage medium can include, but is not limited to, a U disk, a read-only memory, a mobile hard disk, a magnetic disk or an optical disk, and various media capable of storing a computer program.

[0157] In the embodiment 8, an electronic device is disclosed. The electronic device includes a processor and a memory. The memory stores a computer program. The computer program is loaded and executed by the processor to implement a reservoir oil-bearing property evaluation method based on multi-modal data fusion.

[0158] The processor can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA or other programmable logic device, a transistor logic device, a hardware component or any combination thereof. The processor can implement or execute various exemplary logical blocks, modules and circuits described in combination with the disclosure. It can also be a combination of computing functions, such as one or more microprocessor combinations, DSP and microprocessor combinations, etc. The memory can include, but is not limited to, a U disk, a read-only memory, a mobile hard disk, a magnetic disk or an optical disk, and various media capable of storing a computer program.

[0159] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages, such as object-oriented programming language Java and interpreted scripting language JavaScript.

[0160] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the computer or other programmable data processing device produce a device for implementing the functions described in the flowcharts and / or block diagrams. Figure 1 one flow or multiple flows and / or blocksFigure 1 means for performing the function specified by the block or blocks.

[0161] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 Figure 1 means for performing the function specified by the block or blocks.

[0162] The above merely provides the specific implementation of the present application, which has strong adaptability and implementation effect. However, the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the equivalent changes made according to the claims of the present application should be covered within the scope of the present application.​

Claims

1. A method for evaluating reservoir oil-bearing capacity based on multimodal data fusion, characterized in that, include: Establish training sample sets, validation sample sets, and test sample sets. Each sample set includes three categories of oil-bearing samples: high, medium, and low. Each sample includes a rock fragment fluorescence image, rock fragment mineral content data at the corresponding depth, and label information for the oil-bearing evaluation result. The multimodal fusion network is jointly trained end-to-end using the training sample set. Training ends when the training stopping condition is met, resulting in a reservoir oil-bearing evaluation model. The multimodal fusion network is a two-branch weighted fusion structure, with the two branches including an image branch network and a mineral branch network. A self-attention mechanism is used to construct the feature fusion network to achieve dynamic weight allocation. The specific fusion process includes: (1) Feature alignment and projection: The two 512-dimensional feature vectors output by the image branch network and the mineral branch network are L2 normalized to eliminate the feature scale difference between modes in the forward propagation. Then, the projection operation is performed to map the features of the two modes to the shared query. Q ,key K ,value V space: in, This is the weight matrix obtained through training; Features output by the image branching network; Features output by the mineral branching network; (2) Calculate the attention weights. Calculate the correlation AS between the feature vectors output by the image branch network and the feature vectors output by the mineral branch network using the dot product. Then, use the Softmax function to generate the attention weight matrix A: in, This is the transpose of the key matrix; , is the feature dimension scaling factor; (3) Weighted average fusion: The feature vectors output by the image branch network and the feature vectors output by the mineral branch network are weighted and averaged using attention weights to form a fused feature vector. The weight matrix A is automatically updated through backpropagation, with an initial weight of 0.

5. in, To fuse feature vectors; The model parameters of the reservoir oil-bearing evaluation model are optimized based on the validation sample set and the test sample set, and the reservoir oil-bearing evaluation model that meets the test evaluation requirements is output. The reservoir oil-bearing capacity is evaluated using a reservoir oil-bearing capacity evaluation model.

2. The reservoir oil-bearing evaluation method based on multimodal data fusion according to claim 1, characterized in that, Multimodal fusion networks, including: The image branch network extracts features from the rock cuttings fluorescence image to obtain the corresponding 512-dimensional feature vector, which includes a convolutional layer, a max pooling layer, four residual blocks and an adaptation layer. The adaptation layer includes a gradient normalization layer, an adaptive average pooling layer and a fully connected layer. The mineral branching network extracts features from the rock fragment mineral content data to obtain a corresponding 512-dimensional feature vector, which includes three fully connected layers. Each fully connected layer is followed by a gradient normalization layer, a BN layer, and a ReLU activation function. The feature fusion network fuses the 512-dimensional feature vectors output by the image branch network and the 512-dimensional feature vectors output by the mineral branch network to obtain fused features. The classifier, based on fused features, outputs the corresponding reservoir oil-bearing evaluation results.

3. The reservoir oil-bearing evaluation method based on multimodal data fusion according to claim 1 or 2, characterized in that, Establish training sample sets, validation sample sets, and test sample sets, including: Acquire fluorescence images of cuttings at multiple depths in a historical well and determine the mineral content data of the cuttings at the corresponding depths; Based on the fluorescence content threshold, each rock cutting fluorescence image is classified into oil-bearing types to obtain low oil-bearing image label datasets, medium oil-bearing image label datasets, and high oil-bearing image label datasets. Each rock fragment fluorescence image is converted into a corresponding grayscale histogram, and the grayscale histograms in each oil-bearing image label dataset are classified based on histogram correlation. Based on the rules for obtaining oil-bearing sample datasets, we obtained low-oil-bearing sample datasets, medium-oil-bearing sample datasets, and high-oil-bearing sample datasets. The rules for obtaining oil-bearing sample datasets include selecting M gray-level histograms from each of the corresponding gray-level histograms for a given oil-bearing image label dataset, retaining the rock fragment fluorescence images corresponding to each selected gray-level histogram, deleting the remaining rock fragment fluorescence images, and forming M samples. Each sample includes a rock fragment fluorescence image, rock fragment mineral content data at the corresponding depth, and label information for the oil-bearing evaluation result. The low-oil-content sample dataset, medium-oil-content sample dataset, and high-oil-content sample dataset are divided into training set, validation set, and test set according to preset weights, respectively. Sets of the same category are merged to obtain training sample set, validation sample set, and test sample set.

4. The reservoir oil-bearing evaluation method based on multimodal data fusion according to claim 3, characterized in that, It also includes preprocessing of the training sample set, validation sample set, and test sample set, including random pruning, tensor transformation, and Z-score normalization.

5. The reservoir oil-bearing evaluation method based on multimodal data fusion according to claim 1, 2, or 4, characterized in that, Reservoir oil-bearing evaluation is conducted using reservoir oil-bearing evaluation models, including: Obtain fragment fluorescence images and corresponding depth fragment mineral content data of the reservoir to be evaluated; The fluorescence images of rock fragments and the mineral content data of rock fragments at the corresponding depths are input into the reservoir oil-bearing evaluation model to obtain the reservoir oil-bearing evaluation results.

6. A reservoir oil-bearing evaluation device based on multimodal data fusion, applying the method described in any one of claims 1 to 5, characterized in that, include: The sample acquisition unit establishes a training sample set, a validation sample set, and a test sample set. Each sample set contains three types of oil-bearing samples: high, medium, and low. Each sample includes a rock fragment fluorescence image, rock fragment mineral content data at the corresponding depth, and label information for the oil-bearing evaluation result. The model training unit uses the training sample set to perform end-to-end joint training of the multimodal fusion network. When the training stopping condition is met, the training ends and the reservoir oil content evaluation model is obtained. The multimodal fusion network is a two-branch weighted fusion structure, and the two branches include an image branch network and a mineral branch network. The model testing unit optimizes the model parameters of the reservoir oil-bearing evaluation model based on the validation sample set and the test sample set, and outputs a reservoir oil-bearing evaluation model that meets the test evaluation requirements. The reservoir oil-bearing evaluation unit uses the reservoir oil-bearing evaluation model to evaluate the oil-bearing capacity of the reservoir.

7. The reservoir oil-bearing evaluation device based on multimodal data fusion according to claim 6, characterized in that, The sample acquisition unit includes: The basic data acquisition module acquires fluorescence images of cuttings at multiple depths in a historical well and determines the mineral content data of the cuttings at the corresponding depths. The label dataset partitioning module classifies each rock fragment fluorescence image into oil-bearing categories based on fluorescence content thresholds, resulting in low-oil-bearing image label datasets, medium-oil-bearing image label datasets, and high-oil-bearing image label datasets. The grayscale histogram classification module converts each rock cutting fluorescence image into a corresponding grayscale histogram, and classifies the grayscale histograms in each oil-bearing image label dataset based on histogram correlation; The sample dataset acquisition module, based on the rules for acquiring oil-bearing sample datasets, obtains low-oil-bearing sample datasets, medium-oil-bearing sample datasets, and high-oil-bearing sample datasets. The rules for acquiring oil-bearing sample datasets include selecting M gray-level histograms from each of the corresponding gray-level histograms for a given oil-bearing image label dataset, retaining the rock fragment fluorescence images corresponding to each selected gray-level histogram, deleting the remaining rock fragment fluorescence images, and forming M samples. Each sample includes a rock fragment fluorescence image, rock fragment mineral content data at the corresponding depth, and label information for the oil-bearing evaluation result. The sample set partitioning module divides the low oil content sample dataset, medium oil content sample dataset, and high oil content sample dataset into training set, validation set, and test set according to preset weights, and merges sets of the same category to obtain training sample set, validation sample set, and test sample set.

8. The reservoir oil-bearing evaluation device based on multimodal data fusion according to claim 6 or 7, characterized in that, The reservoir oil-bearing evaluation unit includes: The data acquisition module acquires the cuttings fluorescence images and corresponding depth cuttings mineral content data of the reservoir to be evaluated. The oil-bearing evaluation module inputs the rock cutting fluorescence image and the corresponding rock cutting mineral content data at the depth into the reservoir oil-bearing evaluation model to obtain the reservoir oil-bearing evaluation results.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the steps of the method as claimed in any one of claims 1 to 5.

10. A storage medium, characterized in that, The storage medium stores a computer program that can be read by a computer, the computer program being configured to execute the steps of the method as described in any one of claims 1 to 5 when it is run.

Images