Intelligent lithology identification method for water-bearing rock combining image recognition and spectral analysis

Abstract

The application provides a water-bearing rock intelligent lithology identification method combining image recognition and spectrum analysis, comprising the following steps: 1, obtaining original drilling while cutting image and rock infrared spectrum data; 2, performing image pretreatment on the original cutting image; 3, performing cutting sample image feature extraction and normalization processing; 4, performing feature selection on the water-bearing rock spectrum data by using the Filter method and performing normalization processing; 5, based on a double migration feature fusion convolutional neural network, combining the cutting sample image feature and the rock spectrum data to construct a cutting feature analysis model; 6, combining the cutting feature and the rock spectrum data to establish a lithology template library; and 7, performing lithology identification. The application can more accurately identify the lithology in the process of mud logging, and improve the exploration and development efficiency of oil and gas resources.

Description

Technical Field

[0001] This invention relates to the field of oil and gas drilling measurement, and in particular to an intelligent lithology identification method for water-bearing rocks that combines image recognition and spectral analysis. Background Technology

[0002] In the exploration and development of oil and gas resources, cuttings logging is an important logging technology that can intuitively and promptly evaluate the lithology and hydrocarbon potential of underground rocks and establish stratigraphic profiles, providing reliable basic data for geological reservoirs and drilling engineering. The quality of cuttings logging significantly affects the accuracy of oil and gas show identification and reservoir evaluation, and consequently, the subsequent resource development process.

[0003] With the comprehensive development of drilling technology, PDC drill bits, which are characterized by high hardness and good wear resistance, have become a widely used drilling rock-breaking tool. However, due to the high degree of rock fragmentation caused by PDC drill bits, the rock cuttings generated during the drilling process are extremely fine and difficult to identify with the naked eye. The traditional method of manually describing rock cuttings by "observing the color from a distance and examining the lithology up close" is becoming increasingly ineffective, which greatly reduces the efficiency of oil and gas resource exploration. Summary of the Invention

[0004] To address the problems in the background technology, this invention provides an intelligent lithology identification method for water-bearing rocks that combines image recognition and spectral analysis. This method combines image feature recognition with spectral analysis, and utilizes a dual-transfer feature fusion convolutional neural network to construct a comprehensive lithology feature determination model. This enables more accurate lithology identification during well logging, thereby improving the efficiency of oil and gas resource exploration and development.

[0005] Spectroscopic detection technology has been widely used due to its rapid, accurate and non-destructive testing characteristics. Underground water-bearing rocks have different water-bearing characteristics depending on their lithology. By performing spectral analysis on the spectral data of water-bearing rocks collected during drilling, the lithology can be identified in an auxiliary manner during the logging process.

[0006] The specific steps of the intelligent lithology identification method for water-bearing rocks combining image recognition and spectral analysis are as follows:

[0007] Step 1: Acquire raw drilling cuttings images and rock infrared spectral data. The collected cuttings samples are magnified using an optical magnifying glass, then captured by a high-resolution camera, and the images are saved to a computer. Infrared spectral data is obtained by measuring the absorption spectra of the cuttings samples or by acquiring data during drilling.

[0008] Step 2: Preprocess the original rock debris image. Decompose both the low-frequency and high-frequency components of the image using orthogonal wavelet packets of a function set {Um}, m = 0, 1, 2, ..., n. Wavelet packet algorithm: Let... ,then = The corresponding decomposition coefficients are then compared with the threshold and quantized. The decomposition algorithm is as follows:

[0009] pass beg ,

[0010] =

[0011]

[0012] Among them, g n Here, d is the small packet decomposition function, μ is the image mean, a and b are the set decomposition coefficients, t is the image point, and j, n, and k are used to represent the sequence number and the number of decomposition layers.

[0013] Finally, image reconstruction is performed. The reconstruction algorithm is as follows:

[0014] pass , beg

[0015]

[0016] Where h and g are the reconstruction coefficients set, and l and k are used to represent the number of decomposition layers.

[0017] A new image is synthesized using processed coefficients, removing noise from the original image. After wavelet packet denoising, the rock debris image is sharpened using the Laplace operator to make the image clearer and highlight details. After sharpening, the weighting formula is set as follows:

[0018]

[0019] Where Gray represents the image's grayscale value, and R, G, and B represent the three primary colors: red, green, and blue.

[0020] The image is converted to grayscale using the mean method, which facilitates subsequent image feature extraction.

[0021] Step 3: Extract and normalize the features of the rock cuttings sample image. Step 3 extracts texture features based on local grayscale statistics, establishing a... Where L is the image window and L is the gray level. This is a normalized histogram of the window region. The texture features used are as follows:

[0022] Gray mean : ;

[0023] Kuroduku: ;

[0024] Skewness KUP: ;

[0025] Energy Mean Ratio (PMR):

[0026] Contrast CNT: ;

[0027] Histogram entropy : ;

[0028] Histogram invariant moment characteristics :Will Normalization of the k-th order central moments: The characteristic moments of three histograms with invariant displacement and scale were obtained. :

[0029]

[0030] Where M and N are the number of rows and columns of the image, σ is the standard deviation, h(l) is the window area, w determines the horizontal scaling of the function, E is the moment (set manually), and η is the time factor. The normalized value of the k-th order central moment.

[0031] Step 4: Use the Filter method to perform feature selection and normalization on the spectral data of water-bearing rocks. Infrared absorption spectroscopy is based on Beer-Lambert's law:

[0032]

[0033] Absorbance A is inversely proportional to transmittance T and directly proportional to solution thickness b and concentration c. k is the molar absorptivity, which is related to the properties of the absorbing substance and the wavelength λ of the incident light. Underground rocks exhibit certain water-bearing characteristics due to groundwater infiltration. These characteristics vary depending on the lithology. By performing spectral analysis on the infrared spectra of water-bearing rocks, their water-bearing characteristics can be identified and linked to their lithology.

[0034] A filter-type dependency metric was used for feature selection. By analyzing near-infrared spectral data of rocks of different lithologies under different water-bearing characteristics, the relationship between spectral feature variables and their water-bearing characteristics was found, providing basic data for subsequent model building. Six initial feature variables were extracted: peak area, peak height, half-width at half-maximum (HWHM), left shoulder width, right shoulder width, and the ratio of left to right shoulder width. The normalization formula for the feature variables is as follows:

[0035]

[0036] The normalization formula is to subtract the minimum value of the column containing each element from each element of the original data matrix, and then divide by the range of the elements in that column.

[0037] Step 5: Based on the dual transfer feature fusion convolutional neural network, a rock debris feature analysis model is constructed by combining the image features of rock debris samples with rock spectral data. The dual transfer feature fusion learning method is used, involving two feature extraction learning processes: one based on a pre-trained deep transfer learning framework (DDC), and the other a convolutional neural network (CNN) trained directly on the target dataset. These two features are then fused into a new feature, training a new classifier.

[0038] The specific training algorithm flow of this method is as follows:

[0039] Input: Sample dataset With the target dataset Labeled samples from these two datasets.

[0040] Output: Target dataset The category of the sample.

[0041] Ⅰ: In the sample dataset Pre-train a network of k basic source models;

[0042] II: Based on the pre-trained DDC deep transfer learning method, the base source model network obtained in I is applied to the target dataset. Network adaptation learning is performed to obtain k universal feature extractors. ;

[0043] III: Target dataset sample The input is fed into the network model obtained in II, and the output of the general feature extractor is a general feature vector. ;

[0044] IV: Dimensionality reduction is performed on the general features obtained in III to obtain lower-dimensional general features. ;

[0045] V: In the target dataset A special feature extractor is obtained by pre-training a deep CNN model. ;

[0046] VI: Target dataset sample The network model obtained by inputting into II is processed by a special feature extractor to output special feature vectors. ;

[0047] VII: The general feature vector obtained from dimensionality reduction in step III and the specific feature vector from step six are fused with a weight ratio of 6:4 to obtain the sample fused feature vector:

[0048]

[0049] Where β and α are manually set according to the weight ratio, G ft (x) i () is a set of feature vectors obtained by training a model.

[0050] VIII: Obtain the sample fusion features from VII. and its corresponding tags The data is then input into the final classification system for training and prediction.

[0051] Similar to the training process, the prediction algorithm process can be summarized as follows:

[0052] Input: Target dataset Test sample x on;

[0053] Output: The class of sample x;

[0054] Ⅰ: Input the image sample x into the source model after DDC adaptation learning to obtain the general feature vector output by the general feature extractor. Then, based on the projection matrix obtained during training, a low-dimensional general feature vector is obtained after dimensionality reduction. ;

[0055] II: Input sample x into a special CNN model trained on the target dataset to obtain the special feature vector output by the special feature extractor. ;

[0056] III: The general feature vector from step 1 and the special feature vectors in step 2 The fusion is performed to obtain the fused feature vector:

[0057]

[0058] IV: Will The data is input into the final classifier to obtain the category of sample x.

[0059] Step 6: Establish a lithological template library by combining rock fragment characteristics and rock spectral data.

[0060] Steps 3 and 4 are performed on image samples and infrared spectral data of rock fragments of known lithologies to obtain typical features of the lithology. Then, a classifier is obtained by training the features in step 5. A lithology template library is established by combining rock fragment features and rock spectral data.

[0061] Step 7: Perform lithology identification. Feature measurement is achieved using the distance discrimination method, defining the k-th sample in the target rock cuttings sample as... Let the l-th sample in the standard rock cuttings sample be , and the distance between the two types of samples be . , To calculate the average distance between all samples across all categories, the selected feature must satisfy the following formula:

[0062]

[0063]

[0064] in It is the prior probability of class i. For the test set or training set The number of samples in each class. This method uses the s-order Minkowski metric:

[0065] .

[0066] Beneficial effects

[0067] This invention provides an intelligent lithology identification method for water-bearing rocks that combines image recognition and spectral analysis. This method integrates image feature recognition with spectral analysis, employing a dual-transfer feature fusion convolutional neural network to construct a comprehensive lithology feature determination model. Compared to convolutional neural networks used in previous image recognition methods, the dual-transfer feature fusion convolutional neural network method can be trained on smaller datasets, extracts more specific features, and achieves higher recognition accuracy. Compared to simply using image recognition methods for lithology identification, this method, by introducing spectral analysis for auxiliary discrimination and using a dual-transfer feature fusion convolutional neural network to construct a discrimination model, can more accurately identify lithology during well logging, improving the efficiency of oil and gas resource exploration and development. Attached Figure Description

[0068] Figure 1 This is a flowchart of an intelligent lithology identification method for water-bearing rocks that combines image recognition and spectral analysis;

[0069] Figure 2AThis is the original image of the rock debris;

[0070] Figure 2B This is the result of the Laplace operator processing;

[0071] Figure 3A This is the sharpened image;

[0072] Figure 3B This is the final grayscale image after sharpening;

[0073] Figure 4A This is partial infrared spectral data I;

[0074] Figure 4B This is partial infrared spectral data II;

[0075] Figure 4C It is partial infrared spectral data;

[0076] Figure 5 This is a schematic diagram of the basic structure of a convolutional neural network;

[0077] Figure 6 This is the learning process of the double transfer feature fusion convolutional neural network method. Specific Implementation

[0079] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0080] Example

[0081] according to Figure 1-6 The method described above combines image recognition and spectral analysis to identify the type of water-bearing rock.

[0082] Step 1: After magnifying the collected rock cuttings sample with an optical magnifying glass, the image is captured by a high-resolution camera and saved to a computer. Infrared spectral data is obtained by measuring the absorption spectrum of the rock cuttings sample or by collecting data during drilling, thus obtaining the raw rock cuttings images and infrared spectral data required for subsequent processing.

[0083] Step 2: Denoise the image using wavelet packets, then sharpen it using the Laplace operator to make the image clearer and highlight details. Finally, convert the image to grayscale for easier feature extraction later. The original image and the final grayscale image of the rock debris are shown below. Figure 2A , Figure 2B and Figure 3A, Figure 3B As shown.

[0084] Step 3: Extract image features based on local gray-level statistics and perform normalization processing.

[0085] Step 4: Use the Filter method to perform feature selection and normalization on the spectral data of water-bearing rocks. Some infrared spectral data are shown below. Figure 4A , Figure 4B As shown, from Figure 4A , Figure 4B , Figure 4C It can be seen that the spectral curves of rocks of different lithologies have different characteristics.

[0086] Step 5: A new classifier was trained by performing two feature extraction learning operations using a double transfer feature fusion learning method. Based on the double transfer feature fusion convolutional neural network, a rock debris feature analysis model was constructed by combining rock debris sample image features and rock spectral data. The learning processes of the convolutional neural network and this method are as follows: Figure 5 and Figure 6 As shown.

[0087] Step 6: Train the features to obtain a classifier, and establish a lithological template library by combining rock fragment features and rock spectral data.

[0088] Infrared spectroscopy detection and image feature extraction were performed on drilling cuttings samples with calibrated lithology obtained at the well site in the laboratory. Based on the method in step 5, the image and spectral features were trained, and a lithology template library was constructed by combining the calibrated lithology with its image and spectral features.

[0089] Step 7: Implement feature measurement through distance discrimination. Use the s-order Minkowski metric for distance discrimination. The rock type with the smallest distance from the combined sample of rock fragments and infrared spectra is the rock type.

[0090] This method was used to determine the lithology of 100 rock cuttings samples obtained from the well site. The determination results were consistent with the lithology types determined by laboratory calibration. This method can determine the lithology of water-bearing rock cuttings samples.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent lithology identification of water-bearing rocks combining image recognition and spectral analysis, characterized in that: Includes the following steps: Step 1. Acquire raw drilling cuttings images and rock infrared spectral data; After the collected rock cutting particle samples are magnified by an optical magnifying glass, they are photographed by a high-resolution camera and the captured images are saved in a computer. The absorption spectrum of the rock cutting samples is measured, or infrared spectral data is collected during drilling. Step 2. Perform image preprocessing on the original rock cuttings image; Step 3. Extract features from the rock cuttings sample images and perform normalization processing; Step 4. Perform feature selection and normalization on the spectral data of water-bearing rocks; Step 5. Construct a rock cuttings characteristic analysis model; The rock debris feature analysis model was constructed based on a double transfer feature fusion convolutional neural network. When constructing the rock debris feature analysis model, the rock debris sample image features and rock spectral data were combined. The double transfer feature fusion learning method was used to perform two feature extraction learnings. The first was based on a pre-trained deep transfer learning framework, and the second was a convolutional neural network trained directly on the target dataset. The two features were fused into a new feature, and a new classifier was trained. Step 6. Establish a lithological template library by combining rock fragment characteristics and rock spectral data; Step 7. Perform lithological identification.

2. The intelligent lithology identification method for water-bearing rocks combining image recognition and spectral analysis according to claim 1, characterized in that, The image preprocessing method for the original rock debris image in step 2 includes sharpening and converting it to grayscale, and the algorithm is as follows: Let a set of functions be {U m The orthogonal wavelet packet of m=0,1,2...n decomposes both the low-frequency and high-frequency components of the image. The wavelet packet algorithm is as follows: Let... ,then = Then, the corresponding decomposition coefficients are compared with the threshold and quantized. The decomposition algorithm is as follows: pass beg , = where g n is a packet wavelet decomposition function, d is a signal to be decomposed, μ is an image mean, a, b are set decomposition coefficients, t is an image point, j, n, k are used to represent the sequence number and the decomposition level. Finally, image reconstruction is performed. The reconstruction algorithm is as follows: pass , beg Where h and g are the reconstruction coefficients set, and l and k are used to represent the number of decomposition layers; A new image is synthesized by processing the coefficients, removing noise from the original image. After wavelet packet denoising, the rock debris image is sharpened using the Laplace operator to make the image clearer and highlight details. After sharpening, the weighting formula is set as follows: Where Gray represents the image grayscale, and R, G, and B represent the three primary colors of red, green, and blue; Finally, the image is converted to grayscale using the mean method.

3. The intelligent lithology identification method for water-bearing rocks combining image recognition and spectral analysis according to claim 1, characterized in that, The algorithm for extracting and normalizing the features of the rock cuttings sample image in step 3 is as follows: Texture feature extraction is performed based on local grayscale statistics, establishing a... Where L is the image window and L is the gray level. The normalized histogram of the window region uses the following texture features: Gray mean : ; Kuroduku: ; Skewness KUP: ; Energy Mean Ratio (PMR): ; Contrast CNT: ; Rectangular diagram : ; Histogram invariant moment characteristics :Will Normalization of the k-th order central moments: The characteristic moments of three histograms with invariant displacement and scale were obtained. : Where M and N are the number of rows and columns of the image, σ is the standard deviation, h(l) is the window area, w determines the horizontal scaling of the function, E is the moment (set manually), and η is the time factor. The normalized value of the k-th order central moment.

4. The intelligent lithology identification method for water-bearing rocks combining image recognition and spectral analysis according to claim 1, characterized in that, In step 4, feature selection and normalization of the hydrous rock spectral data are performed using the Filter method, the algorithm of which is as follows: Infrared absorption spectroscopy is based on Beer-Lambert's law: Absorbance A is inversely proportional to transmittance T and directly proportional to solution thickness b and concentration c. k is the molar absorptivity, which is related to the properties of the absorbing substance and the wavelength λ of the incident light. Underground rocks have certain water-bearing characteristics under the infiltration of groundwater. Depending on the lithology, their water-bearing characteristics are also different. By performing spectral analysis on the infrared spectrum of water-bearing rocks, the characteristics of their water-bearing features can be identified and linked to their lithology. A filter-type dependency metric was used for feature selection. By analyzing near-infrared spectral data of rocks of different lithologies under different water-bearing characteristics, the relationship between spectral feature variables and their water-bearing characteristics was found, providing basic data for subsequent model building. Six initial feature variables were extracted: peak area, peak height, half-width at half-maximum (HWHM), left shoulder width, right shoulder width, and the ratio of left to right shoulder width. The normalization formula is as follows: The normalization formula means: subtract the minimum value of the column containing each element from each element of the original data matrix, and then divide by the range of the elements in that column.

5. The intelligent lithology identification method for water-bearing rocks combining image recognition and spectral analysis according to claim 1, characterized in that, The lithology identification method in step 7 is the distance discrimination method, which is as follows; Feature measurement is achieved using a distance-based discriminant method, defining the k-th sample in the target rock cuttings sample as... Let the l-th sample in the standard rock cuttings sample be , and the distance between the two types of samples be . , To calculate the average distance between all samples across all categories, the selected feature must satisfy the following formula: in It is the prior probability of class i. For the test set or training set Number of samples in the class n j For the test set or training set V j The number of samples in each class is measured using the s-order Minkowski metric in this method. 。

Images