Automated image-based rock type identification using neural network segmentation and continuous learning

Abstract

A method for evaluating drilling cuttings may include: acquiring a first digital image; and processing the first digital image using a trained neural network (NN) to generate a first segmented image comprising a plurality of tagged segments, wherein at least one tag comprises a lithology type. The segmented image and the acquired first digital image may be processed to retrain the NN. A second digital image is then processed using the retrained NN to generate a second segmented image comprising a plurality of tagged segments, wherein at least one tag comprises a lithology type.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to European Patent Application No. 22305794.4 filed on May 31, 2022, and is incorporated herein by reference in its entirety. Background Art

[0003] In underground drilling operations, such as those used for oil and gas exploration and production, formation cuttings are generated by the drilling activity. These cuttings have long been evaluated and characterized in order to better understand the properties and structure of the subsurface rocks. The cuttings are typically collected after passing through a shale shaker and then cleaned and photographed for analysis. In recent years, there has been a desire to perform this evaluation and characterization automatically (or semi-automatically) to reduce labor costs and shorten the turnaround time for interpretation. However, achieving such automation is far from routine.

[0004] Electronic image data is often manually evaluated during routine drilling operations. Automatic evaluation and characterization of such data is particularly challenging because cuttings include a wide variety of sizes, shapes, textures, and colors. Moreover, cuttings are often jumbled together in the image, with individual particles touching each other and partially overlapping. Image illumination also tends to vary from image to image and equipment to equipment due to shadows and other image artifacts, making evaluation and characterization even more challenging.

[0005] While machine learning algorithms for evaluating rock cuttings images have been disclosed, the machine learning algorithms generally require significant computational time and are often not accurate enough for reliable deployment. There is a need in the industry for methods to improve automated image evaluation to increase the timeliness and accuracy of the evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] For a more complete understanding of the disclosed subject matter and its advantages, reference is now made to the following description in conjunction with the accompanying drawings, in which:

[0007] Figure 1 An exemplary drilling rig including a system for automatically evaluating cuttings images is depicted.

[0008] Figure 2 The depiction is configured from Figure 1 An exemplary device for removing drill bit cuttings from drilling fluid used in the depicted equipment.

[0009] Figure 3 Depicted in Figure 1 An exemplary drill bit that generates drill bit cuttings when used in the depicted apparatus.

[0010] Figure 4 An exemplary process for obtaining rock cuttings images is depicted.

[0011] Figure 5A and 5B (collectively, FIG. 5 ) depicts a flow chart of an exemplary method for segmenting a digital image to identify and label one or more lithology types in the image.

[0012] Figure 6 Graph depicting segmentation accuracy versus depth for an example operation.

[0013] Fig. 7A , 7B 7C, 7D, and 7E (collectively, FIG. 7 ) depict hypothetical examples that further illustrate the disclosed method. DETAILED DESCRIPTION

[0014] The present invention discloses an improved method for segmenting a digital image during a wellbore drilling operation to identify and label one or more lithology types in the image. The disclosed method can advantageously improve the accuracy and timeliness of image evaluation. In an exemplary embodiment, the disclosed method can include: acquiring a first digital image; and processing the first digital image using a trained neural network (NN) to generate a first segmented image including a plurality of labeled segments, wherein at least one label includes a lithology type. The segmented image and the acquired first digital image can be processed to retrain the NN. The retrained NN can then be used to process a second digital image to generate a second segmented image including a plurality of labeled segments, wherein at least one label includes a lithology type.

[0015] As is known to those skilled in the art, drilling of underground wellbores (e.g., in oil and gas exploration and / or production operations) generates a large amount of rock cuttings that can be obtained at the equipment site. In current drilling operations, these cuttings are typically evaluated by geologists, petrophysicists, and / or other experts at the equipment site, for example, to evaluate the local lithology and oil-bearing potential of the well. Cuttings description and evaluation are generally manual and repetitive tasks performed at the equipment site. A commonly implemented workflow involves equipment personnel identifying each lithology type by visually inspecting the cuttings and performing chemical and / or physical measurements on the cuttings samples. After identifying the lithology type in the cuttings sample, the operator immediately describes the rock properties (color, hardness, grain size, etc.) on paper, for example. Photos (digital images) of the cuttings samples can also be visually evaluated based on previous measurements and experience.

[0016] With the increase of image data sets and the progress of machine learning (ML) in recent years, attempts have been made to automatically perform rock cuttings identification using ML algorithms. Such attempts include performing rock cuttings classification from photographic images of drill cuttings (e.g., using deep convolutional neural network analysis). Convolutional neural networks (CNNs) have been trained to perform rock cuttings segmentation so as to classify each pixel, rock or fragment in the rock cuttings image. CNN models have been trained on data sets of rock cuttings sample images collected from lithologies around the world and manually annotated at both the pixel level (assigning labels to each image pixel) and the rock or fragment level (assigning labels to each rock or fragment in the image). Although such efforts are extensive, the efforts have only achieved limited success. A particular difficulty is to accurately label multiple rock types in a single image, such as multiple rock types that may appear during the transition from one lithology to another during drilling operations. A related difficulty is to accurately identify changes from one lithology type to another lithology type in continuous images (this also occurs when transitioning from one lithology type to another lithology type during drilling).

[0017] Figure 1 An exemplary drilling rig 10 including a system 50 configured to automatically evaluate cuttings images is depicted. In the depicted embodiment, a drill string 12 may be suspended at an upper end by a kelly 12 and a traveling block 14, and terminate at a drill bit 16 at a lower end. The drill string 12 and drill bit 16 may be rotated, for example, by a rotary table 18 on a rig floor 20, thereby drilling a borehole (wellbore) 22 in a formation 24, wherein a portion of the borehole 22 may be casing 26. As shown, in certain embodiments, a drilling fluid or drilling "mud" 28 may be pumped into the upper end of the hollow drill string 12 by a mud pump 30 through a connecting mud line 32. From there, the drilling fluid 28 may flow downward through the drill string 12, exit the drill string 12 through an opening in the drill bit 16, and may return to the surface by means of an annulus formed between the wall of the borehole 22 and the outer diameter of the drill string 12. Once at the surface, the drilling fluid 28 may be returned via a return line 34, such as via a bell nipple 36. As shown, in certain embodiments, a blowout preventer 38 may be used to prevent a blowout from occurring during the drilling operation 10.

[0018] like Figure 1 As further depicted, drill bit cuttings formed by the drill bit 16 crushing rock in the formation 24 can be removed from the returning drilling fluid 28 by a shale shaker 40 in a return line 34. The drilling fluid 28 can be reused and recirculated downhole. The shale shaker 40 can include a shaker pit 42 and an air trap 44. Figure 2 Depicted are drill bit cuttings 46 that have been removed from the drilling fluid 28 in the shaker pit 42 of the shale shaker 40 before the drilling fluid 28 is delivered to the mud pit 48. Also, Figure 3Depicted are cuttings generated by the drill bit 16 during drilling and then flowing upward through the wellbore annulus.

[0019] As mentioned above and Figure 1 As depicted, the equipment 10 may include a system 50 configured to automatically evaluate the cuttings image, as described in more detail herein. In certain exemplary embodiments, the system may also include a digital camera configured to take a digital image of the drill cuttings (such as a calibrated digital image). The system 50 may further include computer hardware and software configured to automatically or semi-automatically evaluate the cuttings image. In order to perform these functions, the hardware may include one or more processors (e.g., microprocessors) that may be connected to one or more data storage devices (e.g., a hard drive or solid-state memory to store the processed images). As known to those skilled in the art, the processor may be further connected to a network interface to enable communication with various ground and / or downhole sensors, meters, etc., and to communicate with actuators and other ground and / or downhole equipment. Of course, it should be understood that the disclosed embodiments are not limited to the use or configuration of any particular computer hardware and / or software.

[0020] Now go to Figure 4 , depicting an exemplary process 70 for obtaining a rock fragmentation image. For example, at 72, using the above Figure 1 The wellbore is drilled using the equipment described. The drilling operation also produces rock cuttings as described above. Figure 2 As described, at 74, rock cuttings are collected, for example, from a shale shaker 40. In an exemplary embodiment, the rock cuttings may be prepared for analysis at 76, for example, by drying the rock cuttings in an oven and / or sieving or meshing (e.g., as shown at 90) to remove large particles and / or small particles. The rock cuttings may be further placed on a tray 92 with a high contrast (bright) background color to enhance particle identification and segmentation in the acquired image. The tray 92 may be placed in front of the camera 52, and a photo 94 may be taken at 78. In certain embodiments, the image acquisition process may utilize standardized and / or calibrated lighting, color enhancement, magnification, and / or focus / resolution settings. For example, in certain embodiments, color / illumination calibration is obtained for previously analyzed photos 94 and current photos of interest 94 using a colorimetric algorithm, while resolution calibration may be based on lens focal length, focus distance, and sensor size / resolution so as to compare the current photos of interest 94 with previously analyzed photos 94. Images of the rock cuttings may be taken when they are wet or dry, wherein the humidity is generally controlled for dry rock cutting images.

[0021] Figure 5A and 5B (collectively, FIG. 5 ) depicts a flow chart of exemplary methods 100 and 110 for segmenting a digital image to identify and label one or more lithology types in the image. Figure 5A For example, as mentioned above about Figure 4 As described, at 102, a rock chip image may be obtained. At 104, the received image may be processed using a trained NN (such as a trained CNN) to obtain a segmented image having at least one rock type label. The segmented image may include, for example, a pixel-by-pixel segmentation, where each pixel in the image is assigned a corresponding label (e.g., a lithology type or a background label). At 106, the segmented image obtained at 104 and the acquired image obtained at 102 may be further processed to retrain the NN. For example, as described in more detail below, errors in the segmented image may be corrected (e.g., via relabeling the segmented image) and applied to the original image, which in turn may be used to retrain the NN at 106. Then, at 108, another image (e.g., a subsequent image taken as drilling continues) may be processed using the retrained NN to obtain a segmented image having at least one rock type.

[0022] exist Figure 5B In the embodiment, a rock cutting image may be obtained at 102, and the rock cutting image may be processed at 104 using a trained NN to obtain a segmented image having at least one rock type label (as described above with respect to Figure 5AAt 116, labeling errors may be identified and corrected in the segmented image, thereby further assigning corrected labels (or multiple labels) to certain pixels (or segments) in the acquired image. The identified errors may include, for example, incorrectly labeled rocks in the segmented image (incorrect labels may mean that certain rocks are labeled with incorrect lithology types). At 118, the acquired (and relabeled) image may be divided (cropped) into a plurality of smaller images (crops). These individual crops may advantageously be significantly smaller than the original image. For example, the original image may be divided into a large number of crops, such as in the range of about 20 to about 400 (e.g., 32, 64, 96, 128, 160, 192, 256, or 320) crops per image. At 120, a subset of the crops may be selected and added to a retraining database, which is then used to retrain the NN at 122. In an exemplary embodiment, the subset of selected crops may include 10% or less of the number of crops in the image. Selecting a small number of crops can advantageously improve retraining time, as described in more detail below. In certain exemplary embodiments, when the error identified at 116 exceeds a predetermined threshold, the NN retraining described above with respect to blocks 118, 120, and 122 can be performed. In other embodiments, the NN can be retrained at some preselected interval (e.g., every second, every fourth, every eighth, or every twelfth image). Then, at 124, the retrained NN can be used to process another image (e.g., a subsequent image taken as drilling continues) to obtain a segmented image having at least one rock type label.

[0023] In FIG. 5 , at 104, the acquired images are processed using a trained NN (such as a trained CNN). The trained (or previously trained) NN may have been trained, for example, using a large number of annotated or labeled images, for example, including rock debris images of various lithology types from around the world or from a local basin (oil field or region). The training may result in weight optimization of the NN nodes and layers based on minimizing the loss function of the training data set. In other words, the weights of each node in the multi-layer NN may be tuned so that the model minimizes the error when classifying the input image (or a subset thereof). It should be understood that the NN (e.g., CNN) generally includes multiple layers, and the initial training is intended to train each layer (e.g., each node in each layer is assigned a weight to minimize the loss function). The initial training may require a large amount of computing resources, and therefore may be advantageously performed off-site, for example using an initial training database. Such initial training typically requires several hours (or longer) and is generally not repeated during drilling operations.

[0024] Continuing with reference to FIG. 5 , it will be appreciated that methods 100 and 110 disclose a continuous learning method in which the NN is continuously retrained during drilling operations. By continuously retraining, it is meant that segmentation errors are identified and corrected, and then used to retrain the NN substantially in real time during drilling operations, or at some finite image interval while drilling. Real time may mean that the retraining is completed within a few minutes. Such continuous learning utilizes labeled (or relabeled) images acquired at earlier times or shallower depths in the drilling operation (e.g., at one or more times t-1, t-2, ..., tn or at one or more depths d-1, d-2, ..., dn). Therefore, it will be appreciated that one or more previously acquired images may be used to retrain the NN at 106 and 122.

[0025] In some embodiments, when the segmentation accuracy falls below a predetermined threshold, it may be advantageous to retrain the NN. Figure 6 A graph depicting segmentation accuracy in percentage versus depth (or depth interval) for an exemplary operation. The major lithology types are also labeled on the horizontal axis along with the depth intervals. As depicted, segmentation accuracy is generally very high (e.g., greater than 90%), with occasional drops in accuracy when new lithology types are encountered. It has been found that identifying the lithology type of rock cuttings is very challenging. While NNs are known to be effective at object recognition, the generalization ability of the NNs to new categories is limited or even non-existent. One aspect of the disclosed embodiments is to enable regular (e.g., continuous as defined above) retraining of the NN to enable more effective identification of new lithology types, and rapid improvement in accuracy as drilling continues. In Figure 6 In the figure, the red circle identifies the NN retraining immediately after encountering a new lithology. Note the rapid improvement in accuracy immediately after the retraining.

[0026] Another aspect of the disclosed embodiments is the recognition that it is necessary to balance the retraining data set so that the NN can learn from new lithology types and textures without forgetting what it has previously learned. By way of example only, the NN can be retrained so that it can more effectively recognize shale without forgetting sandstone. This is particularly useful in Figure 690. The model was initially retrained to recognize sandstone at depth 8, and then continued to accurately recognize sandstone cuttings at depths 12 to 72 (where accuracy approached 100%). When shale cuttings were encountered at depth 78, the accuracy dropped significantly. The NN was retrained to recognize shale at depth 78. The accuracy increased significantly at depth 80 as the lithology temporarily transitioned back to sandstone, and then remained high at depth 84 and above as the lithology transitioned back to shale again. By maintaining accuracy across these multiple transitions, the retrained NN demonstrated its ability to recognize a new lithology (shale) while maintaining a memory of the earlier lithology (sandstone).

[0027] It should be appreciated that retraining may involve fine-tuning various NN model parameters, including, for example, the learning rate, the weights given to various lithology classes, and the specific NN layers that may be retrained. Moreover, to enhance memory (or prevent memory loss), the retraining data may include older images (or image crops) of previously encountered lithology types. To further enhance memory and reduce training time, retraining may advantageously retrain only a portion of the NN layers, such as half or less, one-third or less, or one-quarter or less of the NN layers. The retrained layers may further advantageously be the last NN layers. In certain embodiments, retraining may retrain only some of the NN layers, such as the last three or less (e.g., 1, 2, or 3) NN layers. For example, by retraining only a selected subset of the NN layers, retraining time may be reduced to a few minutes, while still significantly improving accuracy.

[0028] Continue to refer Figure 5B , the retraining image may advantageously be divided into subsets of crops (small patches or portions of the original image). A smaller subset of crops may be selected manually or automatically to retrain the model. Using a relatively small subset of crops may advantageously reduce retraining time, which is critical for using method 110 in substantially real time while drilling. It will be appreciated that it is impractical (and extremely expensive) for a drilling operator to wait for the NN to be retrained so that subsequent images can be segmented. To improve retraining accuracy and convergence speed, the crops may be further weighted, shuffled, augmented, transformed (e.g., rotated or flipped), color shifted, and may include erased sections. Such cropping processing may be important in certain operations, such as to reduce overfitting.

[0029] Now turn Figures 7A to 7E (collectively, FIG. 7 ), the disclosed embodiments are described in more detail with the aid of a hypothetical example. As described above, a trained CNN is received or prepared for a drilling operation. In this particular example, the wellbore is initially drilled through a shale formation, and the initial retraining database includes only shale cutouts, such as Fig. 7A As depicted. Figure 7B As depicted, a trained NN was used to process the acquired images. At this particular depth, the formation has changed from shale to primarily sandstone. This resulted in significant labeling errors because the Fig. 7A The retraining database shown is used to retrain and optimize the NN for shale. Figure 7C The software application shown manually re-labels the segmented images.

[0030] Fig.7D Depict an exemplary cropped image. A subset of crops may be selected manually or via the depicted intelligent selection algorithm. The algorithm may advantageously be configured to select only those crops that have the most information about the depicted lithology type (e.g., information related to shape, color, texture, number of individual rocks, etc.). In an exemplary embodiment, the algorithm is configured to select crops that maximize entropy in order to generate an exhaustive and heterogeneous data set. For example, this may be achieved by extracting a feature list for each crop using a NN. A clustering algorithm may then be used to cluster the crops using the extracted feature list and remove duplicate crops (crops that are similar to each other) or crops that are too far (greater than a threshold distance) from the centroid of each cluster.

[0031] Then, if Fig. 7E As depicted, the selected crop can be added to the retraining database. The crops in the retraining database can then optionally be further processed, for example, using the above-described algorithm to further select the most important crops for retraining. Note that the retraining database now includes shale crops and sandstone crops (to improve accuracy and retain memory). Moreover, a weight can be assigned to each crop. For example, increasing the weight can increase the likelihood of selecting a particular crop. In an exemplary embodiment, the more recently an image was marked in the well, the higher the weight it is assigned, thereby increasing the likelihood of selection. From a geological point of view, it should be understood that rocks collected at a depth of 500 meters are more likely to be similar to rocks collected at a depth of 490 meters than to rocks collected at 10 meters. As further depicted (and described above), the retraining database is used to retrain a preselected multi-layer NN. The retrained NN can then be used to process the next rock cutting image. It should be understood that the above-described retraining process can be repeated substantially any suitable number of times while drilling. Depending on the particular stratigraphic structure of the basin, the retraining database may include cutouts with a variety of lithology types.

[0032] It will be understood that the present disclosure includes numerous embodiments. These embodiments include, but are not limited to, the following embodiments.

[0033] In a first embodiment, a method for evaluating drill cuttings obtained during a wellbore drilling operation includes: obtaining a first digital image of the drill cuttings obtained during the wellbore drilling operation; evaluating the first digital image using a trained neural network (NN) to generate a first segmented image including multiple labels, at least one label including a lithology type; correcting labeling errors in the first segmented image and retraining the NN using a portion of the corrected first segmented image; obtaining a second digital image of the drill cuttings obtained during the wellbore drilling operation; and evaluating the second digital image using the retrained NN to generate a second segmented image including multiple labels, at least one label including a lithology type.

[0034] A second embodiment may include the first embodiment, wherein acquiring the first digital image and acquiring the second digital image include: drilling a subterranean wellbore; collecting first drill cuttings and second drill cuttings from a circulating drilling fluid used when drilling the subterranean wellbore; preparing the first drill cuttings and the drilling second cuttings; and taking a first digital photograph of the prepared first drill cuttings and a second digital photograph of the prepared second drill cuttings.

[0035] A third embodiment may include the second embodiment wherein the first drill cuttings are produced at a first depth in the wellbore and the second drill cuttings are produced at a second depth in the wellbore, the second depth being greater than the first depth.

[0036] A fourth embodiment may include any one of the first to third embodiments, wherein correcting the labeling error includes: identifying and correcting the labeling error in the first segmented image to obtain a corrected image; dividing the corrected image into a plurality of crops; selecting a subset of the plurality of crops and adding the selected subset to a retraining database; and retraining the NN using the retraining database.

[0037] A fifth embodiment may include the fourth embodiment, wherein the subset of the plurality of crops is selected to include drill cuttings having a lithology type that was mislabeled in the first segmented image.

[0038] A sixth embodiment may include any one of the fourth to fifth embodiments, wherein the plurality of cutouts includes 20 to 400 cutouts; and the subset includes 10% or less of the cutouts in the plurality of cutouts.

[0039] A seventh embodiment may include any of the fourth to sixth embodiments, wherein the selecting the subset of the plurality of crops is performed automatically to maximize entropy.

[0040] An eighth embodiment may include the seventh embodiment, wherein the selecting the subset of the multiple crops comprises: using the NN to extract a feature list for each of the multiple crops; clustering each of the multiple crops using the extracted feature list; and removing duplicate crops among the clustered crops or crops that exceed a threshold distance from the centroid of each of the clusters to select the subset.

[0041] A ninth embodiment may include any one of the first to eighth embodiments, wherein the NN network includes a plurality of layers, and the retraining of the NN retrains half or fewer of the plurality of layers.

[0042] A tenth embodiment may include the ninth embodiment, wherein the retraining of the NN only retrains the last three layers or less of the NN.

[0043] In an eleventh embodiment, a system for evaluating drill cuttings obtained during a wellbore drilling operation comprises: a digital camera configured to capture digital images of drill cuttings; and a computer processor configured to: cause the digital camera to acquire a first digital image of the drill cuttings obtained during the wellbore drilling operation; utilize a trained neural network (NN) to evaluate the first digital image to generate a first segmented image comprising a plurality of labels, wherein at least one label comprises a lithology type; retrain the NN using a portion of the first segmented image comprising corrected labeling errors; cause the digital camera to acquire a second digital image of the drill cuttings obtained during the wellbore drilling operation; and utilize the retrained NN to evaluate the second digital image to generate a second segmented image comprising a plurality of labels, wherein at least one label comprises a lithology type.

[0044] A twelfth embodiment may include the eleventh embodiment, wherein the retraining of the NN comprises: dividing the first segmented image into a plurality of crops; selecting a subset of the plurality of crops and adding the selected subset to a retraining database; and retraining the NN using the retraining database.

[0045] A thirteenth embodiment may include the twelfth embodiment, wherein the plurality of cutouts includes 20 to 400 cutouts; and the subset includes 10% or less of the cutouts in the plurality of cutouts.

[0046] A fourteenth embodiment may include any one of the twelfth to thirteenth embodiments, wherein the selection of the subset of the multiple crops is automatically performed to maximize entropy and includes: using the NN to extract a feature list for each of the multiple crops; clustering each of the multiple crops using the extracted feature list; and removing duplicate crops among the clustered crops or crops that exceed a threshold distance from the centroid of each cluster in the clusters to select the subset.

[0047] The fifteenth embodiment may include any one of the eleventh to fourteenth embodiments, wherein the NN network includes multiple layers, and the retraining of the NN retrains half or fewer of the multiple layers or only retrains the last three or fewer layers of the NN.

[0048] In a sixteenth embodiment, a method for evaluating drill cuttings obtained during a wellbore drilling operation includes: obtaining a first digital image of the drill cuttings obtained during the wellbore drilling operation; evaluating the first cuttings image using a trained neural network (NN) to generate a first segmented image including multiple labels, at least one of which includes a lithology type; identifying and correcting labeling errors in the first segmented image; dividing the corrected first segmented image into multiple crops; selecting a subset of the multiple crops and adding the selected subset to a retraining database; retraining the NN using the retraining database; obtaining a second digital image of the drill cuttings obtained during the wellbore drilling operation; and evaluating the second digital image using the retrained NN to generate a second segmented image including multiple labels, at least one of which includes a lithology type.

[0049] The seventeenth embodiment may include the sixteenth embodiment, wherein the acquiring the first digital image and the acquiring the second digital image include: drilling an underground wellbore; collecting first drill cuttings and second drill cuttings from a circulating drilling fluid used when drilling the underground wellbore; preparing the first drill cuttings and the drilling second cuttings; and taking a first digital photograph of the prepared first drill cuttings and a second digital photograph of the prepared second drill cuttings.

[0050] An eighteenth embodiment may include any one of the sixteenth to seventeenth embodiments, wherein the crops are selected to include lithology types that are incorrectly labeled in the first segmented image; the multiple crops include 20 to 400 crops; and the subset includes 10% or less of the crops in the multiple crops.

[0051] The nineteenth embodiment may include any one of the sixteenth to eighteenth embodiments, wherein the selection of the subset of the multiple crops is automatically performed to maximize entropy and includes: using the NN to extract a feature list for each of the multiple crops; clustering each of the multiple crops using the extracted feature list; and removing duplicate crops among the clustered crops or crops that exceed a threshold distance from the centroid of each cluster in the clusters to select the subset.

[0052] The twentieth embodiment may include any one of the sixteenth to nineteenth embodiments, wherein the NN network includes multiple layers, and the retraining of the NN retrains half or fewer of the multiple layers or only retrains the last three or fewer layers of the NN.

[0053] Although automated image-based rock type identification with neural network segmentation and continuous learning has been described in detail, it should be understood that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the disclosure as defined by the appended claims.

Claims

1. A method for evaluating drill cuttings obtained during a wellbore drilling operation, the method comprising: acquiring a first digital image of first drill cuttings obtained during a drilling operation in the wellbore; evaluating the first digital image using a trained neural network (NN) to generate a first segmented image comprising a plurality of labels, wherein at least one label comprises a lithology type; correcting labeling errors in the first segmented image and retraining the NN using a portion of the corrected first segmented image; acquiring a second digital image of second drill cuttings obtained during the wellbore drilling operation; as well as The second digital image is evaluated using the retrained NN to generate a second segmented image including a plurality of labels, wherein at least one label includes a lithology type.

2. The method of claim 1, wherein said acquiring the first digital image and said acquiring the second digital image comprises: Drilling underground wellbores; collecting the first drill cuttings and the second drill cuttings from circulating drilling fluid used in drilling the subterranean wellbore; preparing the first drill cuttings and the second drill cuttings; as well as A first digital photograph of the prepared first drill cuttings and a second digital photograph of the prepared second drill cuttings are taken.

3. The method of claim 2, wherein the first drill cuttings are produced at a first depth in the wellbore and the second drill cuttings are produced at a second depth in the wellbore, the second depth being greater than the first depth.

4. The method of claim 1, wherein the correcting the marking error comprises: identifying and correcting labeling errors in the first segmented image to obtain a corrected image; dividing the corrected image into a plurality of crops; selecting a subset of the plurality of crops and adding the selected subset to a retraining database; as well as The NN is retrained using the retraining database. 5 . The method of claim 4 , wherein the subset of the plurality of crops is selected to include drill cuttings having a lithology type that is mislabeled in the first segmented image.

6. The method according to claim 4, wherein: The plurality of cutouts comprises 20 to 400 cutouts; and The subset includes 10% or less of the plurality of crops. The method of claim 4 , wherein said selecting said subset of said plurality of crops is performed automatically to maximize entropy.

8. The method of claim 7, wherein said selecting said subset of said plurality of crops comprises: extracting a feature list for each of the plurality of crops using the NN; Clustering the plurality of cropped parts using the extracted feature list; as well as The subset is selected by removing duplicate crops among the clustered crops or crops that are more than a threshold distance away from the centroid of each of the clusters.

9. The method of claim 1, wherein the NN network comprises a plurality of layers, and the retraining of the NN retrains half or fewer of the plurality of layers.

10. The method of claim 9, wherein the retraining the NN retrains only the last three layers or less of the NN.

11. A system for evaluating drill cuttings obtained during a wellbore drilling operation, the system comprising: a digital camera configured to capture digital images of the drill cuttings; and A computer processor configured to: causing the digital camera to acquire a first digital image of first drill cuttings obtained during a drilling operation in the wellbore; evaluating the first digital image using a trained neural network (NN) to generate a first segmented image comprising a plurality of labels, wherein at least one label comprises a lithology type; retraining the NN using a portion of the first segmented image including the corrected labeling errors; causing the digital camera to acquire a second digital image of second drill cuttings obtained during the wellbore drilling operation; and The second digital image is evaluated using the retrained NN to generate a second segmented image including a plurality of labels, wherein at least one label includes a lithology type.

12. The system of claim 11, wherein said retraining said NN comprises: dividing the first segmented image into a plurality of cropped portions; selecting a subset of the plurality of crops and adding the selected subset to a retraining database; as well as The NN is retrained using the retraining database.

13. The system of claim 12, wherein: The plurality of cutouts comprises 20 to 400 cutouts; and The subset includes 10% or less of the plurality of crops.

14. The system of claim 12, wherein said selecting said subset of said plurality of crops is performed automatically to maximize entropy and comprises: extracting a feature list for each of the plurality of crops using the NN; Clustering the plurality of cropped parts using the extracted feature list; as well as The subset is selected by removing duplicate crops among the clustered crops or crops that are more than a threshold distance away from the centroid of each of the clusters.

15. The system of claim 11, wherein the NN network comprises a plurality of layers, and the retraining of the NN retrains half or fewer of the plurality of layers or retrains only the last three or fewer layers of the NN.

16. A method for evaluating drill cuttings obtained during a wellbore drilling operation, the method comprising: acquiring a first digital image of first drill cuttings obtained during a drilling operation in the wellbore; evaluating the first digital image using a trained neural network (NN) to generate a first segmented image comprising a plurality of labels, wherein at least one label comprises a lithology type; identifying and correcting labeling errors in the first segmented image; dividing the corrected first segmented image into a plurality of crops; selecting a subset of the plurality of crops and adding the selected subset to a retraining database; retraining the NN using the retraining database; acquiring a second digital image of second drill cuttings obtained during the wellbore drilling operation; as well as The second digital image is evaluated using the retrained NN to generate a second segmented image including a plurality of labels, wherein at least one label includes a lithology type.

17. The method of claim 16, wherein said acquiring said first digital image and said acquiring said second digital image comprises: Drilling underground wellbores; collecting the first drill cuttings and the second drill cuttings from circulating drilling fluid used in drilling the subterranean wellbore; preparing the first drill cuttings and the second drill cuttings; as well as A first digital photograph of the prepared first drill cuttings and a second digital photograph of the prepared second drill cuttings are taken.

18. The method of claim 16, wherein: selecting the crop to include lithology types that were mislabeled in the first segmented image; The plurality of cutouts comprises 20 to 400 cutouts; and The subset includes 10% or less of the plurality of crops.

19. The method of claim 16, wherein said selecting said subset of said plurality of crops is performed automatically to maximize entropy and comprises: extracting a feature list for each of the plurality of crops using the NN; Clustering the plurality of cropped parts using the extracted feature list; and removing duplicate clips among the clustered clips or clips whose distance from the centroid of each cluster in the clusters exceeds a threshold value to select the subset.

20. The method of claim 16, wherein the NN network comprises a plurality of layers, and the retraining of the NN retrains half or fewer of the plurality of layers or only retrains the last three or fewer layers of the NN.

Images