Method for the lithological analysis of a sample of drill cuttings

Abstract

The invention relates to a method (50) comprising the steps of: - using a sensor to acquire (110) an image of the sample composed of grains; - applying (120), to the image (1), a convolutional neural network in order to obtain a contour map (C) associating, with each pixel of the image, a value equal to one unit when that pixel belongs to a contour of a grain and otherwise equal to zero; - applying (130), to the contour map (C), a segmentation algorithm in order to obtain a segmented image (S) comprising a plurality of regions, one region being separated from a neighbouring region by a contour having a thickness of a single pixel; and - applying (140), to the segmented image (S), one or more geometric processing operations in order to calculate physical quantities for each grain so as to characterise the lithology of the sample (3).

Description

[0001] TITLE: Lithological analysis method for a drill cuttings sample

[0002] The present invention relates to the field of geological analysis for the optimization of drilling in the energy sector.

[0003] More specifically, the invention relates to the methods of lithological characterization of a sample of drilling debris.

[0004] Drilling cuttings analysis is a crucial step in oil and gas drilling processes. It involves examining samples of drilling mud that rise to the surface during the drilling operation, in order to deduce the geological characteristics of the subsurface.

[0005] These mud samples do indeed contain rock and mineral particles that provide valuable information on the composition and stratigraphy of the subsoil.

[0006] Mud analysis makes it possible to determine the presence of oil, gas or hydrogen reservoirs and to estimate their depth, thickness, as well as the geological and physical properties of the layers traversed.

[0007] It also allows for the assessment of pressure and temperature in the well, which is essential for the safety and design of the facilities.

[0008] Drilling mud analysis is carried out using specific instruments and techniques throughout the drilling process.

[0009] Currently, samples are regularly taken at specific intervals during the drilling operation. These samples are taken directly from the drilling mud that rises to the surface through the well.

[0010] The collected samples are sent to a laboratory where a series of analyses are performed. Commonly used analytical techniques include imaging techniques, mass spectrometry, chromatography, atomic absorption spectroscopy, and other methods to determine the chemical composition of the samples.

[0011] The data obtained from the mud analysis are then interpreted by geologists, drilling engineers and other experts, who use this information to assess the geological composition of the subsoil and deduce the properties of the reservoir, fluid pressure, the presence of contaminants and other important parameters.

[0012] There is a need for an imaging method that can be performed on mud samples that is fast (almost real-time), can be implemented on or in the immediate vicinity of the drilling site, and uses simple and robust hardware, while allowing accurate information to be obtained for lithological characterization of the geological layer being drilled.

[0013] The invention is therefore intended to meet this need.

[0014] To this end, the invention consists of a lithological analysis method for a sample of drilling debris, which may be wet or dry, and comprising a plurality of grains, said method, implemented by a computer associated with a sensor, being characterized in that it comprises, during an operational phase, the steps of: acquisition of an image of the sample by the sensor;application, on the acquired image, of a first deep learning classification algorithm structured as a convolutional neural network, the first classification algorithm being optimized following the implementation of a training phase of the process, so as to obtain a contour map, the contour map associating with each pixel of the acquired image a probability value, which is equal to one when it is estimated that said pixel of the acquired image belongs to a contour of a grain and equal to zero when it is estimated that said pixel of the acquired image does not belong to the contour of a grain; application, on the contour map, of a segmentation algorithm, so as to obtain a segmented image comprising a plurality of regions, a region being separated from a neighboring region by a contour having a thickness of a single pixel, each region being associated with a probable grain delimited in the acquired image;and, application on the segmented image, of one or more geometric processing to calculate a plurality of physical quantities for each probable grain, the values ​​of the calculated physical quantities allowing a lithological characterization of the sample.;

[0015] Depending on specific embodiments, the process comprises one or more of the following characteristics, taken individually or in all technically possible combinations:

[0016] - the first deep learning classification algorithm structured as a convolutional neural network adopts a U-Net structure.

[0017] - the sensor is suitable for acquiring a monochrome or polychrome image of the sample.

[0018] - the sensor is a camera with an extended depth of field allowing imaging of several layers of the sample.

[0019] - the training phase includes an automatic population step of a training database and a training step of the first deep learning classification algorithm structured as a convolutional neural network from the data of the training database, the automatic population step comprising the sub-steps of: acquisition of a training image on a training sample by means of a training sensor identical to the sensor; application, on the acquired training image, of one or more deterministic image processing algorithm(s) allowing to improve a quality of the acquired training image, to obtain a processed training image; application, on the processed training image, of a second deep learning classification algorithm already trained but in a non-specific way, in order to obtain a segmented training image;and, memorization of a new entry in the training database, the new entry associating the acquired training image and the segmented training image estimated for said acquired training image.;

[0020] - the application, on the acquired training image, of one or more deterministic image processing algorithm(s) includes the application of one or more algorithms chosen from: an intensity equalization algorithm; a contrast equalization algorithm; and, a gamma correction algorithm.

[0021] The invention also relates to an analysis system comprising a computer connected to a sensor, the analysis system being adapted for the implementation of a lithological analysis process of a sample of drilling debris conforming to the previous process.

[0022] The invention also relates to a computer program product comprising software instructions which, when executed by a computer, implement the preceding process.

[0023] The invention and its advantages will be better understood upon reading the following detailed description of a particular embodiment, given solely by way of non-limiting example, this description being made with reference to the accompanying drawings in which:

[0024] - Figure 1 is a schematic representation of a system for implementing the analysis method according to the invention;

[0025] - Figure 2 is a block representation of a preferred embodiment of the analysis method according to the invention;

[0026] - Figure 3 is a representation of the different pixel matrices during the implementation of the process shown in Figure 2; and,

[0027] - Figure 4 is a schematic representation of the architecture of the CNN algorithm trained in the training phase and executed in the operational phase of the process of Figure 2.

[0028] Figure 1 is a schematic representation of a system for implementing the analysis method according to the invention. System 1 comprises an optical sensor 2, enabling the acquisition of an image of a sludge sample 3 placed in the field of view of the optical sensor 2.

[0029] System 1 includes a computer 4. The latter includes computing means, such as a processor 5, storage means, such as a memory 6, and an input / output interface 7, enabling in particular the computer 4 to be connected to the sensor 2.

[0030] Memory 6 stores, among other things, computer program instructions, in particular a program 8 whose execution enables the implementation of the method 50 according to the invention.

[0031] The analysis method 50 according to the invention includes a training phase 200 of a parametric model to obtain a trained model using training data from a training database, and then an operational phase 100 using the trained model to perform the analysis of newly acquired images.

[0032] Computer 4 can be used for both the training phase 200 and the operational phase 100. The program 8 then includes on the one hand a training module 12 whose execution, using data from a training database 9, allows the completion of the training phase 200, and on the other hand an operational module 11, whose execution allows the completion of the operational phase 100.

[0033] Alternatively, each phase of the analysis process is carried out on different computer systems, provided that the optical sensors of each of these systems are identical.

[0034] Figure 2 is a schematic block representation of a preferred embodiment of operational phase 100 of analysis process 50.

[0035] Prior to the implementation of operational phase 100, a sample is prepared.

[0036] A drill head, for example of an oil well, regularly brings mud to the surface.

[0037] These muds are characteristic of the geological layer that the drill head has reached and is drilling through.

[0038] Once at the surface, the sludge passes through filter screens to retain only particles (or grains) of a predefined size. For example, the finest and largest grains are removed, leaving only grains of an intermediate size.

[0039] A sample is taken from the outlet of the filter sieves. It therefore consists of an aggregate of grains with weak cohesion. The sample may be wet or dry, after evaporation of all or part of its water content.

[0040] Operational phase 100 begins with a step 110 of acquiring an image I of sample 3.

[0041] The image of the sample is obtained using optical sensor 2.

[0042] Sensor 2 is of the camera type.

[0043] Preferably, it is an improved camera, in the sense that it allows for close-field shooting (macrophotography) of the surface of sample 3, but with an extended depth of field, for example of a few millimeters, preferably between 2 and 3 mm.

[0044] This is advantageous because the sample consists of grains. However, the depth of field in microphotography is shallow, so only grains located precisely in the image plane of the camera are sharp and their edges are clearly defined, while grains located immediately in front of or behind the image plane are blurred and their edges are indistinct. Thus, using a shallow depth of field makes it possible to obtain an image in which all the grains in the sample are sharp and their edges are clearly defined.

[0045] This also allows, when the sample results from the superposition of several layers of grains, to be able to image not only the grains of the top layer, constituting the surface of the sample, but also the layer(s) immediately below the top layer.

[0046] In the preferred embodiment, sensor 2 automatically stacks a series of raw images, obtained from as many different image planes, to extend the depth of field. Sensor 2 thus preferably delivers an image as a preprocessed data file. This is preferable, in terms of time savings, to a sensor delivering the series of raw images and a computer programmed to stack the raw images to obtain the acquired image, which is the subject of the subsequent process.

[0047] The acquired image is a matrix of pixels. Typically, the image contains on the order of 30,000,000 pixels.

[0048] In one embodiment, the acquired image is a chromatic image, for example red-green-blue - RGB (red-green-blue). Each pixel is then associated with three intensity values, respectively a first value for red, a second value for green, and a third value for blue.

[0049] Figure 3A shows a chromatic image. To acquire a chromatic image, the sample is illuminated by white light.

[0050] Alternatively, the acquired image is a monochromatic image, for example in greyscale. Each pixel is then associated with an intensity value.

[0051] For the acquisition of a monochromatic image, the sample is illuminated by monochromatic light, such as ultraviolet - UV light.

[0052] Phase 100 of process 50 then includes a step 120 of analysis of the acquired image I, in order to obtain a contour map C of the sample 3.

[0053] Step 120 consists of applying to the acquired image I a first deep learning classification algorithm structured as a convolutional neural network - CNN (“Convolutional Neural Network”).

[0054] The CNN algorithm, whose structure will be presented in more detail below with reference to Figure 5, is a parametric model. The training phase 200 of the process 50 according to the invention calculates the optimal parameter values ​​to obtain a trained model. This trained model is then used in the operational phase 100 of the process 50.

[0055] The CNN algorithm takes as input the acquired images I at acquisition step 110.

[0056] The CNN algorithm outputs a contour map C. This map has the same pixel resolution as the input image. Each pixel in the map is associated with a unique pixel in the input image. Each pixel in the map is assigned a probability between 0 and 1. A pixel is assigned a value of 1 when the CNN algorithm determines that this pixel belongs to the grain contour in the sample image. A pixel is assigned a value of 0 when the CNN algorithm determines that this pixel does not belong to a grain contour in the sample image.

[0057] It should be noted that on the C contour map, two pixels associated with the value zero can be separated, along the segment that joins them, by several pixels associated with the value unit or close to unit.

[0058] Figure 3B shows a contour map that is estimated from the acquired image shown in Figure 3A.

[0059] Operational phase 100 continues with a step 130 of segmenting the contour mapping C in order to obtain a segmented image S.

[0060] Step 130 involves applying a segmentation algorithm to the C map, grouping pixels according to their neighborhood into a plurality of regions, each region corresponding to a grain of the acquired image. The segmentation algorithm implemented is, for example, of the "watershed" type, familiar to those skilled in the art.

[0061] The result of this segmentation step is, for example, represented by a segmented image, such as image S in Figure 3B.

[0062] The segmented image S is a partition of the acquired image into a plurality of regions, each enclosed by a contour line, with two adjacent regions separated by a contour line only one pixel wide. In other words, the segmented image S is a binary image: the pixels of the different regions are associated with the value zero, while the pixels of the different contours are associated with the value one.

[0063] Segmentation thus makes it possible to very precisely delimit the outline of a probable grain, as identified by the CNN algorithm in the original acquired image.

[0064] Furthermore, each region of the segmented image receives a unique identifier. Step 130 thus allows the pixels of the contour map C to be grouped according to a neighborhood criterion, by associating them with the same identifier.

[0065] Figure 3C shows a segmented image obtained from the contour mapping in Figure 3B. False colors are used to identify each delimited region in the image.

[0066] Operational phase 100 advantageously includes a measurement step 140 from the segmented image S and advantageously also from the acquired image I.

[0067] For each region of the segmented image S, different geometric treatments allow us to measure, for the probable grain corresponding to this region, a set of primary physical quantities, such as its dimensions (surface area, perimeter, average diameter...), its shape (position of the center, principal axis of maximum width, principal axis of minimum width, sphericity...), etc.

[0068] These primary physical quantities are advantageously combined to calculate secondary physical quantities. For example, the roughness of a probable grain can be calculated by the ratio of its perimeter to its surface area.

[0069] The segmented image S is advantageously superimposed, like a mask, on the acquired image I. In the case of an RGB image, for example, different colorimetric processing techniques can then be used to measure a color for each region of the segmented image S, corresponding to the probable grain. The color is an index of the mineral nature of the associated grain.

[0070] Then, the measurements relating to each region are combined to characterize the sample as a whole. For example, a histogram of the grain sphericity measurements of the input image is constructed which, based on the distribution it presents, allows the lithology of the sample to be described.

[0071] For example, a histogram of the surface measurements of the grains in the input image is constructed which, depending on the distribution it presents, allows us to say whether the sample is that of a gravel, a sand, a silt or a clay.

[0072] Preferably, phase 100 is iterated to characterize different samples taken at the same time, i.e., corresponding to the same drilling depth. The accumulation of statistics then allows for a more robust determination of the nature of the geological layer encountered at that drilling depth.

[0073] Advantageously, phase 100 of the process according to the invention is iterated regularly during drilling, so as to obtain geological information for different depths and thus reconstruct the different geological layers traversed by the borehole.

[0074] Since the acquired image analysis step implements an artificial intelligence algorithm, the method according to the invention includes a training phase 200 of the CNN algorithm executed in step 120.

[0075] It should be noted that there is no training database in the field of lithological analysis.

[0076] The training phase 200 of process 50 includes in particular the steps of automatic population of the training database 9 to then allow the implementation of the actual training steps of the CNN algorithm.

[0077] Thus, the training phase 200 of the process 50 includes a step 210 of acquiring a training image.

[0078] This acquisition is carried out under the same conditions as step 110, in particular with the same sensor 2 (or an identical sensor) in order to obtain a training image similar to the acquired image I.

[0079] In the 3D figure, an acquired training image is reproduced.

[0080] Phase 200 continues with a step 215 of pre-processing the training image acquired in step 210. Deterministic image processing is applied to the image to improve its quality and contrast, in order to better bring out the contours of the imaged grains.

[0081] For example, three image processing techniques are implemented.

[0082] The first step is to equalize the brightness of the image. This involves, for example, applying a "histogram equalization" process, which allows the redistribution of the intensity of the pixels in the training image to homogenize the intensity across the entire image.

[0083] Alternatively, one could apply an "adaptive histogram equalization" process, which calculates several histograms, each corresponding to a distinct section of the training image, and uses these different histograms to redistribute the brightness values ​​within the image. This process improves local contrast and edge definition in each section of an image.

[0084] A second processing step is image contrast equalization.

[0085] One example is applying a "Contrast Limited Adaptive Histogram Equalization" (CLAHE) process. This process divides the training image, preferably the output of the first processing, into contextual tiles. It then creates a histogram for each contextual tile, clips it to a predefined value, and redistributes the clipped area to the remaining cells of the histogram. This technique makes hidden features of the training image more visible and sharpens edges in each part of the image.

[0086] The third processing is a gamma correction processing, which allows you to work on the brightness of the training image, preferably at the output of the first or second processing.

[0087] Step 215 produces an IT-processed training image.

[0088] In Figure 3E, a processed training image is reproduced which is derived from the acquired training image reproduced in Figure 3D.

[0089] The next step, 220, of phase 200 is a segmentation step of the processed training image IL

[0090] A second, pre-trained artificial intelligence segmentation algorithm is used to recognize edges in the processed training image. For example, the HED (Holistically-Nested Edge Detection) algorithm is used. This is a convolutional neural network. It is available as open source on the internet and is trained non-specifically on arbitrary images.

[0091] The result of step 220 is a segmented training image Se.

[0092] It should be noted that the execution of the HED algorithm remains lengthy, on the order of 10 minutes to process an image, even a pre-processed one. In Figure 3F, a segmented training image is reproduced, which is estimated from the processed training image in Figure 3E.

[0093] In step 250, a new entry is recorded in the training database 9. The segmented training image Se was used as the label for the training image acquired on.

[0094] Once adequately populated, database 9 is used for training the CNN algorithm in a step 270. The training conforms to the state of the art.

[0095] In operation, the CNN algorithm is capable of producing a contour image S from an acquired image I in a few seconds.

[0096] The process according to the invention is particularly advantageous because it gives very good results not only on dry samples, but also on wet samples.

[0097] The architecture of the CNN algorithm trained in step 270 and executed in step 120 corresponds to the U-Net structure.

[0098] The U-Net structure is described, for example, in the article Ronneberger, O., Fischer, P., & Brox, T. (2015, October). U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention (pp. 234-241). Springer, Cham.

[0099] Thus, the CNN algorithm includes, in series, an encoder 20, a disconnection layer 30, a decoder 40 and an output layer 50.

[0100] The encoder 20 takes as input an acquired image I, for example an RGB image.

[0101] The connection layer 30 is located between the output of the encoder 20 and the input of the decoder 40.

[0102] The decoder 40 outputs a C contour map.

[0103] The coder 20 consists of a plurality of successive convolutional layers, 22j, where i is an index between 1 and N, N denoting the total number of layers of the CNN unit.

[0104] In figure 4, N is equal to 5.

[0105] The information flow passing through encoder 20 goes from the layers with the lowest index (from layer 22-) to the layers with the highest index (towards layer 224).

[0106] Each layer is associated with a specific resolution.

[0107] Each layer 22i typically includes a convolution block to extract a feature map from a reduced feature map produced by the previous layer 22 of the encoder 20. Preferably, the convolution block performs a double convolution, so as to successively apply a first 3x3 convolution, a first ReLu function, a second 3x3 convolution and a second ReLu function to the reduced feature map.

[0108] Layer 22 includes a downscaling block to calculate the reduced feature map for the next layer, from the feature map.

[0109] Preferably, the downscaling block performs a sampling-based discretization process, such as maximum pooling, in particular a 2x2 maximum pooling function.

[0110] The connection layer 30, which links the lower layer 225 of the encoder 20 and the lower layer 425 of the decoder 40, typically includes a convolution block in order to extract a feature map from a reduced-scale feature map emitted by the encoder 20. The reduced-scale feature map is the reduced-scale feature map input to the decoder 40.

[0111] Preferably, the convolution block of the connection layer 30 performs a double convolution, so as to successively apply a first 3x3 convolution, a first ReLu function, a second 3x3 convolution and a second ReLu function on the reduced feature map 36.

[0112] The decoder 40 consists of a plurality of successive convolutional layers, 42, , where i is an index between 1 and N, where N denotes the total number of layers. In the present embodiment, N is equal to 5.

[0113] The information flow passing through decoder 40 goes from the layers with the highest index (from layer 424) to the layers with the lowest index (towards layer 42-,).

[0114] Each layer 42 is associated with a specific resolution, identical to the resolution of layer 22, of the encoder 20 of the same index i.

[0115] Each layer 42 typically includes a scaling block which takes the feature map at the output of the previous layer 42 to determine a scaled feature map.

[0116] Layer 42 also includes a concatenation block performing a concatenation operation between the larger-scale feature map and a refined feature map.

[0117] Layer 42 then includes a transposed convolution block, which performs a transposed convolution on the output of the concatenation block to obtain a feature map. For example, the transposed convolution is a double convolution block that successively applies a first 3x3 convolution, a first ReLu function, a second 3x3 convolution, and a second ReLu function.

[0118] According to the U-Net structure, for a particular resolution, the feature map of the encoder side 20 is transmitted to the decoder side 40 via a jump connection 1 ~l in Figure 4.

[0119] Alternatively, the first AI algorithm in step 130 takes as input several images of the same sample, for example, an RGB image and a UV image. The algorithm must then be trained using training data corresponding to these inputs.

[0120] This method allows for the automatic population of a training database using a trained but non-specific AI algorithm to label images that need to be processed to assist this AI algorithm and thus improve the label quality. This method also allows for the training of an AI algorithm specific to the desired application, in this case, the analysis of sample images.

[0121] The present method allows for near real-time analysis of samples.

Claims

DEMANDS 1. A method (50) for the lithological analysis of a sample (3) of drilling cuttings, which may be wet or dry, and comprising a plurality of grains, said method being implemented by a computer (4) associated with a sensor (2), characterized in that the method comprises, during an operational phase (100), the steps of: - acquisition (110) of an image of the sample by the sensor; - application (120), on the acquired image (I), of a first classification algorithm by deep learning structured as a convolutional neural network, the first classification algorithm being optimized following the implementation of a training phase (200) of the process, so as to obtain a contour map (C), the contour map associating with each pixel of the acquired image a probability value, which is equal to one when it is estimated that said pixel of the acquired image belongs to a contour of a grain and equal to zero when it is estimated that said pixel of the acquired image does not belong to the contour of a grain; - application (130), on the contour map (C), of a segmentation algorithm, so as to obtain a segmented image (S) comprising a plurality of regions, a region being separated from a neighboring region by a contour having a thickness of a single pixel, each region being associated with a probable grain delimited in the acquired image; and, - application (140) on the segmented image (S), of one or more geometric processing to calculate a plurality of physical quantities for each probable grain, the values ​​of the calculated physical quantities allowing a lithological characterization of the sample (3). 2.- Method according to claim 1, wherein the first deep learning classification algorithm structured as a convolutional neural network adopts a U-Net structure. 3.- Method according to claim 1 or claim 2, wherein the sensor (2) is adapted to acquire a monochrome or polychrome image of the sample (3). 4.- Method according to any one of claims 1 to 3, wherein the sensor (2) is a photographic device having an extended depth of field enabling imaging of several layers of the sample.

5. A method according to any one of claims 1 to 4, wherein the training phase (200) comprises a step of automatically populating a training database (9) and a step of learning the first classification algorithm by deep learning structured as a convolutional neural network from the data in the training database, the automatic populating step comprising the substeps of: - acquisition (210) of a training image (the) on a training sample using a training sensor identical to the sensor (2); - application (215), on the acquired training image, of one or more deterministic image processing algorithm(s) allowing to improve a quality of the acquired training image, to obtain a processed training image (It); - application (220), on the processed training image, of a second deep learning classification algorithm already trained but in a non-specific way, in order to obtain a segmented training image (Se); and, - memorization (250) of a new entry in the training database (9), the new entry associating the acquired training image (le) and the segmented training image (Se) estimated for said acquired training image.

6. A method according to claim 5, wherein the application of one or more deterministic image processing algorithms to the acquired training image comprises the application of one or more algorithms chosen from: - an intensity equalization algorithm; - a contrast equalization algorithm; and, - a gamma correction algorithm.

7. Analysis system (1) comprising a computer (4) connected to a sensor (2), characterized in that the analysis system is adapted for the implementation of a lithological analysis process of a drill cutting sample according to any one of claims 1 to 6. 8.- Product computer program comprising program code instructions recorded on a computer-readable medium of a system conforming to claim 7, for the execution of the steps of the lithographic analysis process of a drill cutting sample according to any one of claims 1 to 6 when said program is executed on a computer.

Images