Ultralow-temperature rock microscopic deformation testing system and method based on automatic focusing microscopic DIC

Abstract

The invention belongs to the field of rock physical test devices, and particularly relates to an ultralow-temperature rock microscopic deformation test system and method based on an automatic focusing microscopic DIC. The system comprises a cold and hot table and an optical microscope matched with a CCD camera, an objective table is arranged in the cold and hot table, a glass window is arranged above the cold and hot table, the objective table is connected with a liquid nitrogen tank and a liquid nitrogen pump, an outlet of the liquid nitrogen pump is connected with a liquid nitrogen outlet pipeline, and the liquid nitrogen outlet pipeline points to the glass window; an electric displacement table is arranged below the cold and hot table, a microscope automatic focusing model based on deep learning is arranged in the computer, the model uses an improved lightweight network MobileViTV3, the out-of-focus distance is predicted through a single microscope image obtained on any image plane, and an auxiliary camera or an additional optical device is not needed. The test system and the test method are relatively simple and belong to non-contact strain measurement, a microscope is used for image shooting, and the change process of the rock microscopic pore fracture structure at low temperature can be effectively observed.

Description

Technical Field

[0001] The invention belongs to the field of rock physics test devices, and in particular relates to an ultra-low temperature rock microscopic deformation test system and method based on automatic focusing microscopic DIC (Digital Image Correlation). Background Art

[0002] Compared with traditional fossil energy sources such as coal and oil, natural gas has the advantages of being more efficient and clean. However, due to seasonal changes in LNG demand, uncoordinated import methods, and limitations of storage facilities, the problem of regulating LNG supply and demand is difficult to solve. Under standard atmospheric pressure, the temperature of LNG is below -162°C, and its volume is compressed to one-six-hundred-and-twenty-fifth of its original volume. Therefore, LNG is more conducive to storage and transportation than natural gas. Therefore, it is very important to ensure the safety of large-scale LNG storage facilities and the long-term stable supply management of LNG.

[0003] At present, above-ground storage tanks and underground storage tanks are the two most widely used storage methods in the world. The advantage of above-ground storage tanks is that they are easy to construct and the subsequent update, management, and maintenance costs are relatively low. However, above-ground storage tanks occupy a large area, which is easy to cause waste of land resources, and are greatly affected by the surrounding environment. For example, in the event of bad weather or natural disasters, it is easy to cause the tank to be damaged. Once liquefied natural gas leaks, it is easy to cause major safety problems. The advantages of underground storage tanks are that they are relatively stable, less affected by the environment, and there is a large amount of undeveloped land underground that can effectively save ground space. However, the construction of underground storage tanks is difficult, and the cost of subsequent management and maintenance will be very high.

[0004] At present, mines in many areas are close to the exhaustion stage. Although these abandoned mines have complete infrastructure, they have lost their economic value. However, they have great potential and can be converted into underground natural reservoirs of liquefied natural gas. Compared with ground reservoirs, underground natural reservoirs have a smaller engineering volume, and only the correct site selection is required to balance cost and safety. Therefore, the development of underground natural reservoirs has important practical significance. Since the temperature of liquefied natural gas is very low, when liquefied natural gas comes into contact with rocks, it will cause the water in the rocks to freeze. When the water in the aquifer rocks is frozen, it will produce a volume expansion of about 9%, and the rocks themselves will shrink to a certain extent when they are cooled. This change will have a great impact on the stress state of the rocks. Therefore, studying the freezing and shrinking of rocks at ultra-low temperatures is of great significance to the stability and sealing of underground liquefied natural gas reservoirs.

[0005] The existing methods for measuring the frost heave deformation of rock-like materials are mainly divided into contact measurement and non-contact measurement. The contact measurement method mostly uses the low-temperature strain gauge measurement technology. Measuring rock deformation with low-temperature strain gauges has the advantages of low cost and simple operation. However, the ideal resistance value of the strain gauge for testing frost heave deformation should only change with strain and be unaffected by other factors. In fact, the resistance value of the strain gauge is greatly affected by the ambient temperature (including the test piece). The resistance change caused by the ambient temperature change and the resistance change caused by the strain of the test piece are almost of the same order of magnitude, resulting in a large measurement error. As a non-contact method for measuring low-temperature strain, microscopic DIC can well solve the problem of errors caused by the influence of temperature on the strain gauge. Compared with the strain gauge that can only give the average strain of the measured object, microscopic DIC can also provide the full-field strain distribution information, which is more conducive to researchers to understand the mechanical behavior of materials. Microscopic DIC can also achieve multi-scale deformation analysis by adjusting the magnification of the optical microscope. It can observe and measure the microscopic deformation between rock minerals at the micron scale and also observe and measure the overall deformation of the rock at the macroscopic scale, so as to more comprehensively understand the deformation behavior of the rock at low temperature. Therefore, studying a cryogenic rock mesoscopic deformation test system based on microscopic DIC is of great significance for measuring the frost heave and cold shrinkage characteristics of rocks at ultra-low temperatures and guiding the storage of liquefied natural gas. On this basis, how to achieve automatic focusing of the system is also of great significance for the test accuracy and the convenience of test operation. Summary of the Invention

[0006] In view of the above technical problems, the present invention proposes a cryogenic rock mesoscopic deformation test system, including a heating and cooling stage and a microscope; a specimen stage is arranged inside the heating and cooling stage, and a glass window is provided directly above the specimen stage on the heating and cooling stage; the specimen stage is connected to a liquid nitrogen tank through a second liquid nitrogen pipeline, the specimen stage is connected to the inlet of a liquid nitrogen pump through a first liquid nitrogen pipeline, the outlet of the liquid nitrogen pump is connected to a liquid nitrogen outlet pipeline, and the outlet end of the liquid nitrogen outlet pipeline points to the glass window; an electric displacement stage is arranged under the heating and cooling stage, and the electric displacement stage is sequentially connected to a servo motor and a computer through a first connecting wire, and the computer controls the lifting of the electric displacement stage by controlling the servo motor; the microscope is connected to the computer through a second connecting wire.

[0007] Preferably, it further includes a water circulation pump. A water pipe or a water chamber is arranged on the outer periphery of the heating and cooling stage. One end of the water pipe or the water chamber is connected to the outlet of the water circulation pump through a water inlet pipeline, and the other end is connected to the inlet of the water circulation pump through a water return pipeline.

[0008] Preferably, the inside of the heating and cooling stage is connected to a vacuum pump through an exhaust pipeline.

[0009] Preferably, the microscope is an optical achromatic microscope equipped with a CCD camera.

[0010] Preferably, the temperature controller is connected to the liquid nitrogen pump to control the operation of the liquid nitrogen pump.

[0011] Preferably, a temperature sensor is arranged in the stage.

[0012] The present invention provides a method for constructing an automatic focusing model of a microscope based on deep learning, comprising the following steps:

[0013] S1. Use different objective lenses equipped on the microscope to continuously photograph the observation object from the far focus to the near focus at a certain step length, capture a series of images on different focal planes, and convert the captured images into grayscale images. These grayscale images form a Z-Stack.

[0014] S2. Use the Brenner Gradient algorithm to process the grayscale images, find the grayscale image with the largest Brenner Gradient value, and this grayscale image is identified as the clearest grayscale image and used as the ideal focal plane image. Among them, the Brenner Gradient value is defined as:

[0015]

[0016] where H represents the height of the grayscale image, corresponding to the x-axis, W represents the width of the grayscale image, corresponding to the y-axis, and f(x, y) represents the grayscale value of the coordinate point corresponding to the grayscale image.

[0017] S3. For each grayscale image in the Z-Stack, assign a label value according to its distance from the ideal focal plane image, and the focal distance label of the ideal focal plane image is zero.

[0018] S4. Crop each grayscale image into multiple grayscale image blocks to obtain a dataset.

[0019] S5. Construct an automatic focusing model of the microscope based on deep learning. The model includes a two-dimensional convolutional layer, a feature extraction layer, a two-dimensional convolutional layer, a global average pooling layer, and a fully connected layer connected in sequence. The first two-dimensional convolutional layer includes a convolutional layer and downsampling. A single grayscale image block F∈R h×w , is used as the input of the model, where h is the height of a single grayscale image block, w is the width of a single grayscale image block, and R represents a two-dimensional matrix or array.

[0020] The feature extraction layer includes 5 layers. The first layer is 1 MV2 block; the second layer is 2 MV2 blocks, and the first MV2 block has downsampling; the structures of the third to fifth layers are the same, including MV2 blocks and MobileViT blocks, and the MV2 blocks have downsampling. After processing, a single grayscale feature map is obtained where h1 is the height of a single grayscale feature map, w 1 is the width of a single grayscale feature map, and c is the number of channels;

[0021] The second convolutional layer only includes a two-dimensional convolutional layer;

[0022] The global average pooling first calculates the average value d of all elements within each channel of the grayscale feature map i , and then combines the average value d of all elements within each channel i into a vector d of length c, d ∈ R c×1 , d i The expression of is:

[0023]

[0024] where, F ijk represents the i-th channel, height h of a single grayscale feature map 1 at the j-th row, width w 1 at the k-th column of the element value;

[0025] The fully connected layer includes a weight matrix W ∈ R 1×c and a bias vector b ∈ R 1×1 , and the fully connected layer is used to convert the output d ∈ R of the global average pooling c×1 into the final output result z, that is, the defocus distance of the grayscale image block, which is also the focal distance label. The output value z of the fully connected layer is calculated by the following linear transformation: z = Vd + b;

[0026] S6. Use the Adam optimizer and the smooth L1 loss function to train the model;

[0027] S7. Model training: Use the training set data to train the model, calculate the prediction result through forward propagation, update the model parameters through backward propagation. After each epoch ends, use the validation set to evaluate the model performance and adjust the hyperparameters.

[0028] Preferably, in step S4, data augmentation is performed on the obtained grayscale image blocks to obtain a grayscale image block dataset, and the data augmentation method includes image rotation and flipping.

[0029] Preferably, in step S4, the dataset is divided into a training set, a validation set, and a test set.

[0030] Step S7 also includes: Model testing: Use the test set to finally evaluate the model to evaluate the generalization ability of the model; Model saving: If the model evaluation is qualified, save the model.

[0031] The present invention also provides a microscope autofocus model based on deep learning, which is constructed by using the method for constructing a microscope autofocus model based on deep learning described above.

[0032] The present invention also provides a cryogenic mesoscopic deformation test system for rocks based on autofocus microscopy DIC, including the cryogenic mesoscopic deformation test system described above. The computer of the cryogenic mesoscopic deformation test system for rocks contains a computer-readable storage medium, and the computer-readable storage medium stores a computer program. When the computer program is executed, the method for constructing a microscope autofocus model based on deep learning described above is realized.

[0033] The present invention also provides a cryogenic mesoscopic deformation test system for rocks based on autofocus microscopy DIC, including the cryogenic mesoscopic deformation test system described above. The computer of the cryogenic mesoscopic deformation test system for rocks is provided with the microscope autofocus model based on deep learning described above.

[0034] The present invention also provides a cryogenic mesoscopic deformation test method for rocks based on autofocus microscopy DIC, using the cryogenic mesoscopic deformation test system for rocks based on autofocus microscopy DIC described above; specifically including the following steps:

[0035] Step 1: Grind and dry the upper and lower surfaces of the rock sample to obtain a completely dry rock sample.

[0036] Step 2: Place it in a vacuum container and seal it, add deionized water until a completely saturated rock sample is obtained; dry it again, and obtain rock samples with different masses, that is, different saturations, by controlling the drying time.

[0037] Step 3: Spray a layer of white paint on the surface of the rock sample. After waiting for the white paint to dry completely, use a high-pressure spray gun to spray an alcohol-silica powder solution on the surface of the rock sample. Different-sized speckles can be obtained by controlling the different mesh numbers of the silica powder.

[0038] Step 4: Place the rock sample in the middle of the stage and align it vertically with the glass window and the eyepiece of the microscope.

[0039] Step 5: Use the microscope autofocus model based on deep learning for autofocus.

[0040] Step 6: Open the liquid nitrogen pump. Nitrogen in the liquid nitrogen tank enters the stage and then flows out from the liquid nitrogen outlet pipeline through the liquid nitrogen pump. Control the temperature of the rock sample and perform different-rate cooling treatments on the stage, and take pictures of the rock sample after each cooling for recording.

[0041] Step 7: Use the 2D VIC-2D software to analyze and process the collected rock sample images.

[0042] Preferably, the fifth step specifically includes:

[0043] a. Obtain an image: Capture an image of the current rock sample through a microscope and convert it into a grayscale image;

[0044] b. Clarity judgment: Use the Brenner Gradient algorithm to process the grayscale image of the rock sample, compare the obtained Brenner Gradient value with a preset value, and judge whether its clarity reaches the preset standard; if so, execute step f; if not, execute step c;

[0045] c. Predict the defocus distance: Crop the grayscale image of the rock sample into multiple grayscale image blocks; use a microscope autofocus model based on deep learning to predict the defocus distance of each grayscale image block, and take the average value of the defocus distances of all grayscale image blocks as the defocus distance of the current rock sample;

[0046] d. Adjust the focal length: Control the servo motor through a computer, and then control the electric displacement stage to move according to the defocus distance;

[0047] e. Iterative adjustment: Repeat steps a - d until the clarity determined in step b reaches the preset standard;

[0048] f. Maintain the focal length and complete the autofocus adjustment process.

[0049] The beneficial technical effects of the present invention are as follows:

[0050] 1. The test system of the present invention can effectively measure the deformation characteristics of rocks under ultra - low temperature conditions. With the help of digital image processing technology, this system can synchronously obtain strain data of rocks in multiple directions during a single test process, thus providing a more efficient and accurate measurement means for the study of the anisotropic deformation behavior of rocks.

[0051] 2. The test system and test method of the present invention are relatively simple and belong to non - contact strain measurement, which makes up for the problems of low - temperature failure of traditional strain gauges and complex operation.

[0052] 3. The test system of the present invention uses a microscope for image capture, which can effectively observe the change process of the microscopic pore - fracture structure of rocks at low temperatures.

[0053] 4. Compared with other two - dimensional DIC tests, the present invention introduces Z - axis correction, effectively solves the error caused by the change in sample height, improves the test accuracy, and is more conducive to observing the tiny deformation of rocks caused by temperature factors.

[0054] 5. The present invention predicts the defocus distance through a single image, reducing the focusing time and improving the efficiency of autofocus; the autofocus model of the present invention uses a lightweight network with low computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The drawings constituting the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0056] Figure 1 It is a schematic diagram of the overall system of the ultra-low temperature rock mesoscopic deformation test system of the present invention;

[0057] Figure 2 It is a schematic diagram of the structure of the autofocus microscopy DIC model based on deep learning of the present invention;

[0058] Figure 3 It is a schematic diagram of the autofocus process based on the autofocus microscopy DIC model of the present invention;

[0059] Description of reference numerals: 1 is a microscope; 2 is a hot and cold stage; 3 is a temperature controller; 4 is a liquid nitrogen pump; 5 is a water circulation pump; 6 is a liquid nitrogen tank; 7 is a vacuum pump; 8 is a computer; 9 is an electric displacement table; 10 is a servo motor; 11 is a water inlet pipeline; 12 is a water return pipeline; 13 is a first liquid nitrogen pipeline; 14 is a liquid nitrogen outlet pipeline; 15 is a second liquid nitrogen pipeline; 16 is an exhaust pipeline; 17 is a first connection line; 18 is a second connection line; 19 is a third connection line; 20 is a fourth connection line; 21 is a glass window; 22 is a stage. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] The following will describe in detail the specific embodiments of the present invention with reference to the drawings.

[0061] Embodiment 1

[0062] As Figure 1As shown in the figure, the ultra-low temperature rock mesoscopic deformation test system of the present invention includes a heating and cooling stage 2 and a microscope 1; a stage 22 is arranged inside the heating and cooling stage 2, and a glass window 21 is located directly above the stage 22 on the heating and cooling stage 2; the stage 22 is connected to a liquid nitrogen tank 6 through a second liquid nitrogen pipeline 15, the stage 22 is connected to the inlet of a liquid nitrogen pump 4 through a first liquid nitrogen pipeline 13, the outlet of the liquid nitrogen pump 4 is connected to a liquid nitrogen outlet pipeline 14, and the outlet end of the liquid nitrogen outlet pipeline 14 points to the glass window 21. The stage 22 is a cavity structure; it further includes a water circulation pump 5. A water pipe or a water cavity is arranged on the outer periphery of the heating and cooling stage 2. One end of the water pipe or the water cavity is connected to the outlet of the water circulation pump 5 through a water inlet pipeline 11, and one end is connected to the inlet of the water circulation pump 6 through a water return pipeline 12. The water pipe or the water cavity can increase the temperature of the heating and cooling stage 2 (neutralize the low temperature generated by a part of the liquid nitrogen) to prevent damage to some temperature-intolerant components; the inside of the heating and cooling stage is connected to a vacuum pump 7 through an exhaust pipeline 16 to prevent frosting; an electric displacement stage 9 is arranged under the heating and cooling stage 2. The electric displacement stage 9 is sequentially connected to a servo motor 10 and a computer 8 through a first connecting line 17. The servo motor 10 is used to drive the lifting of the electric displacement stage 9. The computer 10 controls the lifting and the lifting amount of the electric displacement stage 9 by controlling the servo motor 10; the microscope 1 selects an Olympus BX53 microscope. The microscope 1 is connected to the computer 8 through a second connecting line 18. The computer 8 is used to control the operation of the microscope 1; the computer 8 is further connected to a temperature controller 3 through a third connecting line 19. The temperature controller 3 is connected to the liquid nitrogen pump 4 to control the operation of the liquid nitrogen pump 4, and further control the temperature of the test sample on the stage 22; a temperature sensor (not shown in the figure), such as a 100Ω platinum RTD, can be arranged in the stage 22, and a PID temperature control method is adopted.

[0063] Embodiment 2

[0064] As Figure 2 shown in the figure, the present invention also proposes a method for constructing an automatic focusing model of a microscope based on deep learning, which can achieve single-shot automatic focusing. The specific steps are as follows:

[0065] S1. Use different objective lenses equipped on the microscope 1 to continuously photograph the observation object from the far focus to the near focus at a certain step (for example, 3μm), capture a series of images on different focal planes, and convert the captured images into grayscale images. These grayscale images constitute a Z-Stack;

[0066] S2. Process the grayscale image using the Brenner Gradient algorithm to find the grayscale image with the largest Brenner Gradient value, which is identified as the clearest grayscale image and used as the ideal focal plane image. The Brenner Gradient value is defined as:

[0067]

[0068] where H represents the height of the grayscale image, corresponding to the x-axis, W represents the width of the grayscale image, corresponding to the y-axis, and f(x, y) represents the grayscale value of the corresponding coordinate point in the grayscale image.

[0069] S3. For each grayscale image in the Z-Stack, assign a label value according to its distance from the ideal focal plane image, and the focal distance label of the ideal focal plane image is zero.

[0070] S4. Crop each grayscale image into multiple grayscale image patches of 672×672 pixels. To improve the generalization ability of the model, perform data augmentation on the obtained grayscale image patches to obtain a grayscale image patch dataset. The data augmentation methods include image rotation and flipping, etc. Divide the dataset into a training set, a validation set, and a test set, and allocate them according to 8:1:1.

[0071] S5. As Figure 2 shown, construct a microscope autofocus model based on deep learning. The model includes a two-dimensional convolutional layer, a feature extraction layer, a two-dimensional convolutional layer, a global average pooling layer, and a fully connected layer connected in sequence. The first two-dimensional convolutional layer uses a 3×3 convolutional layer, which is used for feature extraction, expanding the input channels, and has downsampling with a downsampling factor of 2 to reduce the spatial resolution of the data to reduce the model complexity. A single grayscale image patch F∈R h×w , as the input of the model, where h is the height of a single grayscale image patch, w is the width of a single grayscale image patch, and R represents a two-dimensional matrix or array.

[0072] The feature extraction layer includes 5 layers. The first layer is 1 MV2 block. The second layer is 2 MV2 blocks, and the first MV2 block has downsampling with a downsampling factor of 2. The structures of the third to fifth layers are the same, including MV2 blocks and MobileViT blocks, and the MV2 blocks have downsampling with a downsampling factor of 2. The MV2 block is an inverted residual structure from MobileNetV2, and the MobileViT block is the core component of the MobileViTV3 model. It extracts grayscale image patch features by fusing local and global features, and at the same time has downsampling with a downsampling factor of 2 to reduce the spatial resolution of the data to reduce the model complexity. After processing, a single grayscale feature map is obtained where h 1= h / 32, w 1 = w / 32, where c is the number of channels, equivalent to c h 1 × w 1 two-dimensional matrix or array;

[0073] The second two-dimensional convolutional layer uses a 1×1 convolutional layer to compress the number of channels of the grayscale feature map;

[0074] The global average pooling layer is used to map the grayscale feature map to a fixed size for subsequent fully connected layer processing. Specifically, the global average pooling first calculates the average value d of all elements in each channel of the grayscale feature map i , and then combines the average value d of all elements in each channel i into a vector d of length c (d ∈ R c×1 ), and the expression of d i is:

[0075]

[0076] where h 1 is the height of a single grayscale feature map, w 1 is the width of a single grayscale feature map, F ijk represents the element value at the i-th channel, the j-th row at height h and the k-th column at width w 1 of a single grayscale feature map 1 ;

[0077] The parameters of the fully connected layer include a weight matrix V ∈ R 1×c and a bias vector b ∈ R 1×1 , and the fully connected layer is used to convert the output d ∈ R c×1 of the global average pooling into the final output result z, which is the defocus distance of the grayscale image block and also the focal length distance label. The output value z of the fully connected layer is calculated through the following linear transformation: z = Vd + b;

[0078] S6. The Adam optimizer and the Smooth L1 Loss function are used to train the model. The mathematical definition of the Smooth L1 Loss function is:

[0079]

[0080] where a represents the difference between the predicted value and the target value;

[0081] S7. Model training: Use the training set data to train the model, calculate the prediction results through forward propagation, update the model parameters through backward propagation. After each epoch (completely traversing the training set once), use the validation set to evaluate the model performance and adjust the hyperparameters; Model testing: Use the test set to finally evaluate the model to evaluate the generalization ability of the model; Model saving: If the model evaluation is qualified, save the model.

[0082] The present invention also claims a microscope autofocus model based on deep learning, which is constructed by using the method for constructing a microscope autofocus model based on deep learning described above.

[0083] Embodiment III

[0084] The present invention also claims a cryogenic rock meso-deformation test system based on autofocus microscopy DIC, including the cryogenic rock meso-deformation test system shown in Embodiment I. The computer of the cryogenic rock meso-deformation test system based on autofocus microscopy DIC contains the microscope autofocus model based on deep learning in Embodiment II, or the computer of the cryogenic rock meso-deformation test system based on autofocus microscopy DIC contains a computer-readable storage medium, and the computer-readable storage medium stores a computer program. When the computer program is executed, it realizes the method for constructing a microscope autofocus model based on deep learning in Embodiment II.

[0085] Embodiment IV

[0086] As Figure 3 shown, the present invention also proposes a cryogenic rock meso-deformation test method based on autofocus microscopy DIC. Conduct tests based on the test system in Embodiment I, use the microscope autofocus model based on deep learning constructed in Embodiment II for focusing. The constructed microscope autofocus model based on deep learning is set in the computer 8, or operate the method for constructing a microscope autofocus model based on deep learning in Embodiment II in the computer 8 to obtain a microscope autofocus model based on deep learning; or conduct tests using the cryogenic rock meso-deformation test system based on autofocus microscopy DIC in Embodiment III;

[0087] The cryogenic rock meso-deformation test specifically includes the following steps:

[0088] First step: Polish the upper and lower surfaces of the rock sample with 400, 800, 1200, and 2000-mesh sandpapers in sequence to obtain a rock sample with flat upper and lower surfaces; Place the rock sample in a drying oven and dry it at 110 °C to obtain a completely dry rock sample;

[0089] Step 2: Place the completely dried rock sample into a vacuum container, seal it, evacuate the air until a vacuum state is reached, add deionized water to the vacuum container to submerge the rock sample, keep the vacuum container in the vacuum state, take out the rock sample and weigh it until a completely saturated rock sample is obtained; put the rock sample into the drying oven again, different masses of rock samples can be obtained by controlling the drying time, wrap the obtained rock samples with plastic wrap, and let them stand until the moisture inside the rock samples diffuses evenly; calculate the corresponding saturation according to the sample mass;

[0090] Step 3: Place the rock sample horizontally, evenly spray a layer of white paint on the surface of the rock sample, let it stand for 30 minutes, and after the white paint is completely dry, use a high-pressure spray gun to spray an alcohol-silica powder solution on the surface of the rock sample. Different-sized speckles can be obtained by controlling the different mesh numbers of the silica powder;

[0091] Step 4: Place the rock sample in the middle of the stage 22, and align it vertically with the glass window 21 and the eyepiece of the microscope 1; turn on the vacuum pump 7 to evacuate the inside of the heating and cooling stage 2;

[0092] Step 5: Perform autofocus using a microscope autofocus model based on deep learning, including

[0093] a. Obtain an image: Capture the current rock sample image through the microscope 1 and convert it into a grayscale image;

[0094] b. Clarity judgment: Use the Brenner Gradient algorithm to process the grayscale image of the rock sample, compare the obtained Brenner Gradient value with a preset value, and judge whether its clarity reaches the preset standard; if so, execute step f; if not, execute step c;

[0095] c. Predict the defocus distance: Crop the grayscale image of the rock sample into multiple grayscale image blocks of 672×672 pixels; use a microscope autofocus model based on deep learning to predict the defocus distance of each grayscale image block, and take the average value of the defocus distances of all grayscale image blocks as the defocus distance of the current rock sample;

[0096] d. Adjust the focal length: Control the servo motor 10 through the computer 8, and then control the electric displacement stage 9 to move according to the defocus distance;

[0097] e. Iterative adjustment: Repeat steps a-d until the clarity determined in step b reaches the preset standard;

[0098] f. Keep the focal length and complete the autofocus adjustment process;

[0099] Step 6: Turn on the liquid nitrogen pump 4 and the water circulation pump 5. Nitrogen in the liquid nitrogen tank 6 enters the stage 22, and then flows out from the liquid nitrogen outlet pipeline 14 through the liquid nitrogen pump 4. The temperature of the rock sample is controlled between -190 °C and 5 °C by operating the temperature controller 3 through a computer, and the stage 22 is cooled at different rates. The cooled rock sample is photographed and recorded every 5 °C through the microscope 1;

[0100] Step 7: Use the 2D VIC-2D software to analyze and process the collected images of the rock samples.

[0101] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to listing the above embodiments. It should be noted that all equivalent replacements and obvious deformation forms made by any person skilled in the art under the guidance of this specification fall within the substantial scope of this specification and should be protected by the present invention.

Claims

1. An ultra-low temperature rock micro-deformation test system, comprising a hot and cold stage and a microscope; characterized in that: A stage is arranged inside the hot and cold table, and a glass window is arranged above the hot and cold table opposite to the stage; the stage is connected to a liquid nitrogen tank through a second liquid nitrogen pipeline, the stage is connected to an inlet of a liquid nitrogen pump through a first liquid nitrogen pipeline, the outlet of the liquid nitrogen pump is connected to a liquid nitrogen outlet pipeline, and the outlet end of the liquid nitrogen outlet pipeline points to the glass window; an electric displacement stage is arranged under the hot and cold table, the electric displacement stage is connected to a servo motor and a computer in sequence through a first connecting line, and the computer controls the lifting and lowering of the electric displacement stage by controlling the servo motor; the microscope is connected to the computer through a second connecting line.

2. The ultra-low temperature rock micro-deformation test system according to claim 1 is characterized in that: It also includes a water circulation pump, a water pipe or a water cavity is arranged on the periphery of the hot and cold stage, one end of the water pipe or the water cavity is connected to the water outlet of the water circulation pump through a water inlet pipeline, and the other end is connected to the water inlet of the water circulation pump through a return pipeline; and / or, the inside of the hot and cold stage is connected to a vacuum pump through an exhaust pipeline; and / or, the microscope is an optical achromatic microscope equipped with a CCD camera; and / or, a temperature controller is connected to a liquid nitrogen pump to control the operation of the liquid nitrogen pump; and / or, a temperature sensor is arranged in the stage.

3. A method for constructing a microscope autofocus model based on deep learning, characterized in that: The steps include: S1. Use a microscope equipped with different objective lenses to continuously shoot the observed object from far focus to near focus in a certain step length, capture a series of images at different focal planes, and convert the captured images into grayscale images. These grayscale images constitute a Z-Stack; S2. Use the Brenner Gradient algorithm to process the grayscale image and find the grayscale image with the largest Brenner Gradient value. This grayscale image is identified as the clearest grayscale image and used as the ideal focal plane image. The Brenner Gradient value is defined as: Where H represents the height of the grayscale image, corresponding to the x-axis, W represents the width of the grayscale image, corresponding to the y-axis, and f(x,y) represents the grayscale value of the corresponding coordinate point of the grayscale image; S3. For each grayscale image in the Z-Stack, a label value is assigned according to its distance from the ideal focal plane image, and the focal distance label of the ideal focal plane image is zero; S4. Crop each grayscale image into multiple grayscale image blocks to obtain a data set; S5. Construct a microscope autofocus model based on deep learning, the model includes a two-dimensional convolution layer, a feature extraction layer, a two-dimensional convolution layer, a global average pooling layer and a fully connected layer connected in sequence; the first two-dimensional convolution layer includes a convolution layer and downsampling; a single grayscale image block F∈R h×w , as the input of the model, where h is the height of a single grayscale image block, w is the width of a single grayscale image block, and R represents a two-dimensional matrix or array; The feature extraction layer includes 5 layers, the first layer is 1 MV2 block; the second layer is 2 MV2 blocks, and the first MV2 block is downsampled; the third to fifth layers have the same structure, including MV2 blocks and MobileViT blocks, and the MV2 blocks are downsampled; after processing, a single grayscale feature map is obtained Among them, h1 is the height of a single grayscale feature map, w1 is the width of a single grayscale feature map, and c is the number of channels; The second convolutional layer only includes a 2D convolutional layer; The global average pooling first calculates the average value d of all elements in each channel of the grayscale feature map i , and then take the average value d of all elements in each channel i Combined into a vector d of length c, d∈R c×1 , d i The expression is: Among them, F ijk Represents a single grayscale feature map The element value of the i-th channel, the j-th row at height h1, and the k-th column at width w1; The fully connected layer consists of a weight matrix V∈R 1×c and a bias vector b∈R 1×1 , the fully connected layer is used to average the output of the global pooling d∈R c×1 Converted to the final output result z, that is, the defocus distance of the grayscale image block, which is also the focal distance label, the output value z of the fully connected layer is calculated by the following linear transformation: z = Vd + b; S6. Use Adam optimizer and smooth L1 loss function to train the model; S7. Model training: Use the training set data to train the model, calculate the prediction results through forward propagation, update the model parameters through back propagation, and use the validation set to evaluate the model performance and adjust the hyperparameters after each epoch.

4. The method for constructing a microscope autofocus model based on deep learning according to claim 3, characterized in that: In step S4, data enhancement is performed on the obtained grayscale image blocks to obtain a grayscale image block data set, and the data enhancement method includes image rotation and flipping; And / or, in step S4, dividing the data set into a training set, a validation set and a test set; And / or, step S7 further includes: model testing: using the test set to perform a final evaluation on the model to evaluate the generalization ability of the model; Model saving: If the model is qualified, save the model.

5. A microscope auto-focusing model based on deep learning, constructed using the microscope auto-focusing model construction method based on deep learning described in claim 3 or 4.

6. Ultra-low temperature rock microscopic deformation testing system based on automatic focusing micro-DIC, characterized by: It comprises the ultra-low temperature rock mesoscopic deformation test system as described in claim 1 or 2, wherein the computer of the ultra-low temperature rock mesoscopic deformation test system comprises a computer-readable storage medium, and the computer-readable storage medium stores a computer program, and when the computer program is executed, the method for constructing a microscope automatic focusing model based on deep learning as described in claim 3 or 4 is implemented.

7. Ultra-low temperature rock microscopic deformation testing system based on automatic focusing micro-DIC, characterized by: It comprises the ultra-low temperature rock mesoscopic deformation test system as described in claim 1 or 2, wherein the computer of the ultra-low temperature rock mesoscopic deformation test system is provided with the deep learning-based microscope automatic focusing model as described in claim 5.

8. A method for testing ultra-low temperature rock microscopic deformation based on auto-focusing microscopic DIC, using the ultra-low temperature rock microscopic deformation testing system based on auto-focusing microscopic DIC according to claim 6 or 7, characterized in that: The following steps are involved: Step 1: Grind the upper and lower surfaces of the rock sample and dry it to obtain a completely dry rock sample; Step 2: Place the sample in a vacuum container and seal it, add deionized water until a fully saturated rock sample is obtained; dry it again, and obtain rock samples of different qualities, i.e., different saturations, by controlling the drying time; Step 3: Spray a layer of white paint on the surface of the rock sample. After the white paint is completely dry, use a high-pressure spray gun to spray alcohol-silicon powder solution on the surface of the rock sample. By controlling the mesh size of the silicon powder, different sizes of speckles can be obtained. Step 4: Place the rock sample in the middle of the stage and align it vertically with the glass window and the eyepiece of the microscope; Step 5: Use the deep learning-based microscope autofocus model for autofocusing; Step 6: Turn on the liquid nitrogen pump, and the nitrogen in the liquid nitrogen tank enters the stage, and then flows out from the liquid nitrogen outlet pipeline through the liquid nitrogen pump to control the temperature of the rock sample. The stage is cooled at different rates, and the rock sample is photographed and recorded after each cooling; Step 7: Use two-dimensional VIC-2D software to analyze and process the collected rock sample images.

9. The ultra-low temperature rock microscopic deformation testing method based on automatic focusing according to claim 8 is characterized in that: The fifth step specifically includes: a. Image acquisition: Capture the current rock sample image through a microscope and convert it into a grayscale image; b. Clarity judgment: Use the Brenner Gradient algorithm to process the grayscale image of the rock sample, compare the obtained Brenner Gradient value with the preset value, and judge whether its clarity meets the preset standard; if yes, execute step f; if not, execute step c; c. Predicting the defocus distance: Crop the grayscale image of the rock sample into multiple grayscale image blocks; use the microscope autofocus model based on deep learning to predict the defocus distance of each grayscale image block, and take the average of the defocus distances of all grayscale image blocks, which is the defocus distance of the current rock sample; d. Adjust the focal length: Control the servo motor through the computer, and then control the electric translation stage to move according to the defocus distance; e. Iterative adjustment: repeat steps ad until the clarity determined in step b reaches the preset standard; f. Maintain the focus and complete the automatic focus adjustment process.

Images