Fluorescence automatic logging system based on deep learning and detection method thereof

Abstract

The application discloses a kind of fluorescence automatic mud logging system and detection method based on deep learning, system is by industrial camera, purple light tube, mobile terminal, cloud server composition, the industrial camera is used to photograph rock sample image, the light tube is used to irradiate rock sample, the mobile terminal is used to control industrial camera, purple light tube and send and receive image data, the cloud server is used to realize rock sample image processing and analysis.Its effect is: one side through the semantic segmentation technology of deep learning image processing, detect rock fluorescence color and brightness, carry out fluorescence quantitative analysis;The other side will be deployed in cloud server deep learning algorithm, it is convenient that fluorescence mud logging personnel can carry out fluorescence mud logging work at any time through mobile device, ensure algorithm analysis efficiency and accuracy, improve work efficiency.

Description

Technical Field

[0001] This invention relates to the fields of petroleum exploration and image processing, semantic segmentation, and deep learning, specifically to a deep learning-based fluorescence automatic logging system and its detection method. Background Technology

[0002] Since the development of logging technology, a series of modern integrated logging methods have been formed, including cuttings logging, gas logging, and fluorescence logging. Among them, fluorescence logging is the simplest, most intuitive, and practical effective method for initially locating oil and gas show zones in petroleum exploration. Fluorescence logging technology is still a commonly used method for detecting oil during drilling. It involves illuminating samples with a fluorescent lamp and directly observing the fluorescence occurrence, color, and intensity with the naked eye to determine whether oil is present and its grade (Wu Shenghe, Cai Zhengqi, Shi Shangming. Oil Geology. Beijing: Petroleum Industry Press, 2011: 22-69). This method can only provide qualitative interpretation and has significant limitations.

[0003] Therefore, there is an urgent need to develop a quantitative fluorescence logging system to better solve some problems such as fluorescence that cannot be identified by the naked eye and is affected by human factors in conventional fluorescence logging, so as to achieve quantitative interpretation and be more conducive to the discovery and identification of light oil with fluorescence wavelengths below 400nm. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies by proposing a deep learning-based automated fluorescence logging system and its detection method. This system applies semantic segmentation technology from deep learning image processing to fluorescence logging, meeting the requirements for quantitative fluorescence logging. Simultaneously, it fully utilizes the convenience brought by artificial intelligence, reducing the consumption of human resources.

[0005] To achieve the above objectives, the specific technical solution adopted by the present invention is as follows:

[0006] This invention provides a deep learning-based fluorescence automatic logging system and its detection method. The key features are: a camera module, a fluorescent lamp, a mobile terminal, and a cloud server. The camera module is used to capture images of rock cuttings samples, the fluorescent lamp is used to illuminate the rock cuttings samples, the mobile terminal is used to control the industrial camera and the ultraviolet lamp, and to send and receive image data, and the cloud server is used to process and analyze the rock cuttings sample images.

[0007] This invention first provides a deep learning-based fluorescence automatic logging system, which consists of a camera module, a UV lamp, a mobile terminal, and a cloud server. The camera module is used to capture images of rock cuttings samples, the UV lamp is used to illuminate the rock cuttings samples, the mobile terminal is used to control the camera module and the UV lamp, as well as to send and receive image data, and the cloud server is used to process and analyze the rock cuttings sample images.

[0008] Preferably, the camera module has adjustment functions for exposure, gain, saturation, and sharpness parameters. Specifically, the camera module uses a Medvision MV-SUA1600C industrial camera.

[0009] In one embodiment, the ultraviolet lamp is a 365nm wavelength ultraviolet lamp.

[0010] Preferably, the mobile terminal is a mobile phone or tablet computer, which communicates with the camera module wirelessly or via wired connection, such as USB. The operator can adjust the camera module's on / off state and camera parameters on the mobile terminal, such as exposure time, analog gain, saturation, and sharpness.

[0011] In one embodiment, the cloud server has a built-in deep learning-based fluorescence detection algorithm that performs fluorescence detection by processing and analyzing images. Preferably, the deep learning-based fluorescence detection algorithm uses the Deeplabv3 algorithm.

[0012] More specifically, the Deeplabv3 algorithm is as follows: in the original Deeplabv3 network structure, ResNet is used as the backbone feature extraction network, and an efficient channel attention module and an attention category feature module are added.

[0013] The efficient channel attention module works as follows: First, a global average pooling operation is used on the effective feature layer A to obtain the feature vector A1. Then, the feature vector A1 is convolved using a one-dimensional convolution kernel of size k, and then passed through a sigmoid activation function to obtain the final feature vector A2. Finally, the corresponding elements of the effective feature layer A and the feature vector A2 are multiplied to obtain the final effective feature layer B. Here, k is adaptively determined through the mapping of the channel dimension c, expressed by the formula:

[0014]

[0015] Where γ = 2, b = 1;

[0016] Specifically, the attention category feature module calculates the category on the effective feature layer M after feature extraction, which is built upon the backbone feature extraction network incorporating an efficient channel attention module. The average probability of pixels Then calculate the results The feature layer is obtained by multiplying the probability values ​​of all pixels belonging to the category, and then the feature layer is connected to the effective feature layer M. Finally, the final effective features are obtained by enhancing feature extraction through the ASPP structure.

[0017] As a further description, the improved Deeplabv3 algorithm deployed on the cloud server has undergone the following improvements:

[0018] ① In the original Deeplabv3 network structure, ResNet is used as the backbone feature extraction network, and an efficient channel attention module (ECANet) is added. First, global average pooling is used on the effective feature layer A to obtain feature vector A1. Then, feature vector A1 is convolved with a one-dimensional convolution kernel of size k, and then passed through a sigmoid activation function to obtain the final feature vector A2. Finally, the effective feature layer A and feature vector A2 are multiplied at corresponding positions to obtain the final effective feature layer B; where k is adaptively determined through the mapping of the channel dimension c, expressed by the formula: Where γ = 2 and b = 1. This structure effectively enables cross-channel interaction, assisting the base network in improving segmentation performance and enhancing the accuracy of the original model;

[0019] ② In the original Deeplabv3 network structure, after the backbone feature extraction network completes feature extraction, an attention category feature module (ACFNet) is added. That is, on the effective feature layer M after feature extraction in the ResNet backbone feature extraction network with the addition of the efficient channel attention module, the category is calculated. The average probability of pixels Then calculate the results The feature layer is obtained by multiplying the probability values ​​of all pixels belonging to the category, connecting the feature layer to the effective feature layer M, and finally performing enhanced feature extraction through the ASPP structure.

[0020] Finally, the output of the neural network was processed, and the image pixel data was divided into three categories: background (non-rock debris samples), rock debris samples without obvious fluorescence, and rock debris samples with obvious fluorescence. The selected detection index was the proportion of rock debris samples with obvious fluorescence among all rock debris samples, calculated using the following formula 1 to reflect the fluorescence level of the rock debris sample:

[0021]

[0022] Where, N w×h N is the total number of pixels in the image. f N is the total number of pixels in the rock debris sample that produces obvious fluorescence. b This represents the total number of pixels in the non-rock debris sample.

[0023] In conjunction with the above system, this invention also provides a deep learning-based method for detecting rock cutting fluorescence, which mainly includes the following steps:

[0024] S1: When performing fluorescence detection, the mobile terminal will acquire rock debris images captured by the camera module via wired or wireless means, such as USB, or acquire rock debris images from the mobile terminal's local album.

[0025] S2: Send the rock debris images to the cloud server via the network.

[0026] S3: Analyze rock debris sample areas and generate prediction results using the improved Deeplabv3 algorithm deployed on a cloud server;

[0027] S4: The server returns the prediction results and sends them to the mobile terminal for display.

[0028] This invention, building upon existing fluorescence logging technology, fully utilizes artificial intelligence. On one hand, it employs semantic segmentation technology in deep learning image processing to detect the fluorescence color and brightness of rock cuttings, enabling quantitative fluorescence analysis. On the other hand, it deploys deep learning algorithms on cloud servers, allowing fluorescence logging personnel to perform their work anytime via mobile devices, ensuring algorithmic analysis efficiency and accuracy, and improving work efficiency. Therefore, the intelligent quantitative fluorescence logging system provided by this invention effectively solves some problems associated with conventional fluorescence logging, such as fluorescence that is undetectable by the naked eye and is affected by human factors, achieving quantitative interpretation and being more conducive to the discovery and identification of light oil with fluorescence wavelengths below 400nm. Attached Figure Description

[0029] Figure 1 This is a system principle block diagram of the present invention.

[0030] Figure 2 This is a flowchart of the detection process of the present invention.

[0031] Figure 3 This is a diagram of the main network structure of the improved Deeplabv3 according to the present invention.

[0032] Figure 4 This is a diagram of the improved Deeplabv3 network Bottleneck1 structure of this invention.

[0033] Figure 5 This is a diagram of the improved Deeplabv3 network Bottleneck2 structure of this invention.

[0034] Figure 6 This is a structural diagram of the ECANet high-efficiency channel attention module of the present invention.

[0035] Figure 7 This is a diagram of the ACFNet class feature attention network structure of the present invention.

[0036] Figure 8 This is a result diagram of an application embodiment of the present invention. Detailed Implementation

[0037] The specific embodiments and working principles of the present invention will be further described in detail below with reference to the accompanying drawings.

[0038] like Figure 1 As shown, a deep learning-based automatic fluorescence logging system and its detection method are disclosed. The deep learning-based automatic fluorescence logging system comprises a camera module, a fluorescent lamp, a mobile terminal, and a cloud server. The camera module is used to capture images of rock cuttings samples, the fluorescent lamp is used to illuminate the rock cuttings samples, the mobile terminal is used to control the camera module and the fluorescent lamp, and the server is used to process and analyze the rock cuttings sample images.

[0039] In the specific implementation process, the camera module uses a Medvision MV-SUA1600C industrial camera with a resolution of 4608×3456, and can adjust parameters such as exposure time, analog gain, saturation, and sharpness. The fluorescent lamp uses a 365nm wavelength ultraviolet lamp. The mobile terminal is a tablet computer. The detection method of the deep learning-based automatic fluorescence logging system, in the specific implementation process, such as... Figure 2 As shown, the operator needs to choose between using a camera to capture an image or using an image from the photo album within the fluorescence detection system. When using the camera, the mobile terminal acquires and displays the video stream captured by the camera module. After the operator clicks the photo button, the detection system acquires the captured image and uploads it to the cloud server, where it waits for the cloud server to analyze the image and return the results. When using an image from the photo album, the detection system automatically opens the mobile terminal's local photo album, acquires the image selected by the operator, uploads it to the cloud server, and waits for the cloud server to analyze the image and return the results.

[0040] In the control process of the above system, whether images are captured by a camera or images are from a photo album, the image analysis process includes the following steps:

[0041] S1: Send the rock debris images to the cloud server via the network.

[0042] S2: Analyze the rock debris sample area and generate prediction results using the improved Deeplabv3 algorithm deployed on a cloud server;

[0043] S3: The server returns the prediction results and sends them to the mobile terminal for display.

[0044] In the above process, the improved Deeplabv3 algorithm deployed on the cloud server mainly performs convolution and pooling operations on the digital rock cuttings image to extract deep features. During encoding, features are extracted from the initial effective feature layer after four compressions and then merged. During decoding, the number of channels in the initial effective feature layer is adjusted, and then stacked with the upsampled results of the effective feature layer after dilated convolution. Finally, the stacked features are used to obtain the prediction result. This classifies each pixel in the digital rock cuttings image and obtains the fluorescence percentage. The Deeplabv3 network was improved as follows:

[0045] ① In the original Deeplabv3 network structure, ResNet is used as the backbone feature extraction network, and an efficient channel attention module (ECANet) is added. Specifically, first, global average pooling is used on the effective feature layer A to obtain feature vector A1. Then, feature vector A1 is convolved with a one-dimensional convolution kernel of size k, and then passed through a sigmoid activation function to obtain the final feature vector A2. Finally, the effective feature layer A and feature vector A2 are multiplied at corresponding positions to obtain the final effective feature layer B; where k is adaptively determined through the mapping of the channel dimension c, expressed by the formula: Where γ = 2 and b = 1. This structure effectively enables cross-channel interaction, assisting the base network in improving segmentation performance and enhancing the accuracy of the original model;

[0046] ② In the original Deeplabv3 network structure, after the backbone feature extraction network completes feature extraction, an attention category feature module (ACFNet) is added. That is, on the effective feature layer M after the ResNet backbone feature extraction network completes feature extraction, the average probability of pixels belonging to a category is calculated. Then calculate the results The layer is obtained by multiplying the probability values ​​of all pixels belonging to the category, and then connected as a new feature to the effective feature layer M. Finally, the feature extraction is enhanced through the ASPP structure.

[0047] Finally, the output of the neural network is processed, and the proportion of fluorescence in the rock fragments within the entire rock fragment sample is calculated using Formula 1. This reflects the fluorescence intensity of the rock fragment sample.

[0048]

[0049] Where, N w×h N is the total number of pixels in the image. f N is the total number of pixels in the rock debris sample that produces obvious fluorescence. bThis represents the total number of pixels in the non-rock debris samples. The image pixel data is divided into three categories: the background of the non-rock debris samples, rock debris samples that do not produce obvious fluorescence, and rock debris samples that produce obvious fluorescence.

[0050] Application Example 1

[0051] In a practical cuttings fluorescence detection experiment, a mobile terminal (specifically a Lenovo Xiaoxin Pad Pro) was used to run an automated fluorescence logging system, acquiring digital image information of the cuttings sample numbered WZ11-1N-A6S1 via a camera module. The specific implementation process is as follows: Figure 2 As shown, the detailed process is as follows:

[0052] ① After preparing the WZ11-1N-A6S1 cuttings sample, run the fluorescence automated logging system using a mobile terminal. The system will allow the operator to choose between using the camera module to capture images or using images from the local photo album.

[0053] ② This implementation selects the camera module to capture images. The camera module receives instructions from the mobile terminal to capture images of the WZ11-1N-A6S1 rock cuttings and sends the images to the cloud server.

[0054] ③ The cloud server will preprocess the received WZ11-1N-A6S1 rock debris images. The captured image resolution of 4608×3456 is not suitable for deep neural network processing, so the image is first reshaped to 4608×3072, and then the image is cut into 54 parts, that is, each part of the image has a resolution of 512×512.

[0055] ④ Each image is input into the Deeplabv3 deep learning semantic segmentation network, and after calculation by the deep neural network, the classification of each pixel in each image is finally obtained, including: rock debris samples, rock debris fluorescence, non-rock debris samples and fluorescence.

[0056] ⑤ Map the Deeplabv3 network output of each image to the image, and then stitch the 54 mapped images back together to form the original image. Classify all pixels and calculate the fluorescence percentage using Formula 1.

[0057] ⑥ Return the results to the mobile terminal for display.

[0058] In the above embodiments, the analysis results of the fluorescence automatic logging system are as follows: Figure 8As shown in the image, the left image is the original image captured by the camera module, where the rock debris sample is laid out in a circular shape, with scattered fluorescent spots. The right image is the inference result of the fluorescence detection algorithm (dark black represents the background, dark gray represents the rock debris sample, and light gray represents the fluorescence of the rock debris), clearly marking the outline of the rock debris sample and the fluorescent areas. The calculation results of the fluorescence detection algorithm are displayed in the lower left corner of the right image, including: the total number of sample pixels, the total number of fluorescent pixels in the rock debris, and the fluorescence percentage calculated using Formula 1 (3.07%). This indicates that the fluorescence level of this sample is relatively low. Figure 8 It can be seen that the fluorescent spots marked by the fluorescence detection algorithm are basically consistent with the human eye's judgment, indicating that the algorithm's calculation results are basically correct.

Claims

1. A fluorescence automatic logging system based on deep learning, characterized in that: It consists of a camera module, a UV lamp, a mobile terminal, and a cloud server. The camera module is used to capture images of rock cuttings samples, the UV lamp is used to illuminate the rock cuttings samples, the mobile terminal is used to control the camera module and the UV lamp, as well as to send and receive image data, and the cloud server is used to process and analyze rock cuttings sample images. The cloud server has a built-in fluorescence detection algorithm based on deep learning, which performs fluorescence detection by processing and analyzing images; The fluorescence detection algorithm based on deep learning adopts the Deeplabv3 algorithm: in the original Deeplabv3 network structure, ResNet is used as the backbone feature extraction network, and an efficient channel attention module and an attention category feature module are added. The efficient channel attention module works as follows: First, global average pooling is used on the effective feature layer A to obtain feature vector A1. Then, feature vector A1 is convolved using a one-dimensional convolution kernel of size k, and then passed through a sigmoid activation function to obtain the final feature vector A2. Finally, the corresponding elements of the effective feature layer A and feature vector A2 are multiplied to obtain the final effective feature layer B. Here, k is adaptively determined through the mapping of the channel dimension c, expressed by the formula: Where γ = 2, b = 1; Specifically, the attention category feature module calculates the average probability of pixels belonging to category i on the effective feature layer M after feature extraction, which is built upon the backbone feature extraction network incorporating an efficient channel attention module. Then calculate the results The feature layer is obtained by multiplying the probability values ​​of all pixels belonging to that category. and feature layer The features are connected to the effective feature layer M, and finally the final effective features are obtained by enhancing feature extraction through the ASPP structure. Finally, the output of the neural network is processed, and the fluorescence percentage is calculated using Formula 1, reflecting the fluorescence level of the rock cutting sample: Where, N w×h N is the total number of pixels in the image. f N is the total number of pixels in the rock debris sample that produces obvious fluorescence. b It represents the total number of pixels in the non-rock debris samples; among them, the image pixel data is divided into three categories: the background of the non-rock debris samples, rock debris samples that do not produce obvious fluorescence, and rock debris samples that produce obvious fluorescence.

2. The deep learning-based fluorescence automatic logging system according to claim 1, characterized in that: The camera module has adjustment functions for exposure, gain, saturation, and sharpness parameters.

3. The deep learning-based fluorescence automatic logging system according to claim 1, characterized in that: The ultraviolet lamp tube uses a 365nm wavelength ultraviolet lamp tube.

4. The deep learning-based fluorescence automatic logging system according to claim 1, characterized in that: The mobile terminal is a mobile phone or tablet computer, which communicates with the camera module wirelessly or via wired means.

5. The deep learning-based fluorescence automatic logging system according to claim 1, characterized in that: The wired connection is via USB.

6. The detection method for fluorescence automatic logging using the deep learning-based fluorescence automatic logging system as described in any one of claims 1 to 5.

7. The detection method as described in claim 6, characterized in that, Includes the following steps: S1: When performing fluorescence detection, the mobile terminal acquires rock debris images captured by the camera module via USB, or the mobile terminal directly acquires rock debris images from its local photo album. S2: Send the rock cuttings images to the cloud server via the network; S3: Analyze rock debris sample areas and generate prediction results using the improved Deeplabv3 algorithm deployed on a cloud server; S4: The cloud server returns the prediction results and sends them to the mobile terminal for display.

Images