A method and device for preventing drill pipe sticking based on multi-modal semantics and a storage medium

Abstract

This invention discloses a multimodal semantic-based anti-sticking control method, device, and storage medium, belonging to the field of intelligent control technology for oil and gas exploration. Addressing the problems of multi-target adhesion, feature artifacts, and time delay misalignment between the surface and bottom of the well in existing logging, this invention deploys the acquisition platform in the sand discharge drop zone of the vibrating screen, constructing a gravity-discrete field of view with a 15° to 30° airflow shield to achieve spatial decoupling of physical morphology; it maps real-time operating conditions into a dynamic cue word matrix, injecting it as a weighted penalty factor into the cross-attention layer of the large visual model; it filters overlapping pseudo-boundaries by calculating local curvature gradients to obtain the sharpness of the true morphology; finally, it calculates the time delay phase difference between visual features and surface mechanical torque by combining the annular cuttings return time. This invention eliminates the spatiotemporal lag of visual detection, directly outputting electrical signals to the drilling rig bus, achieving a leap from passive data monitoring to active closed-loop defense of the underlying physical equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment control technology for oil and gas exploration and development, and more specifically, to a method, device, and storage medium for preventing stuck drill pipe based on multimodal semantics. Background Technology

[0002] In well logging operations, the grain size distribution (PSD) of cuttings and abnormal rock fragments are key indicators for assessing wellbore stability, wellbore cleaning efficiency, and preventing stuck pipe. Existing visual recognition methods for cuttings from vibrating screens typically follow a process of "acquiring images of the screen surface -> traditional convolutional neural network (CNN) object detection -> size calculation".

[0003] However, existing technologies suffer from the following insurmountable technical defects when dealing with complex real-world drilling conditions, preventing them from being truly applicable to industrial-grade closed-loop control:

[0004] First, the physical morphology is severely coupled, resulting in high computational redundancy. Existing image acquisition equipment is positioned directly opposite the vibrating screen, which is not only subject to interference from the 50Hz high-frequency vibration, but also to the overlapping and obstruction of rock debris on the screen surface, making it very easy to generate "perimeter artifacts" at the algorithm layer, leading to distortion of the extracted morphological sharpness.

[0005] Second, the model lacks geological semantics and cross-attention constraints. Traditional visual models are trained offline using static labels and rely on simple pixel-level features, which cannot distinguish between "large rock fragments generated by normal drilling" and "abnormal rockfalls caused by formation stress collapse". They are prone to failure when faced with different geological strata and fluid environments.

[0006] Third, a fatal "spatiotemporal misalignment" leads to control lag. Existing systems only display passively measured visual results, ignoring the significant time delay in the return of cuttings from the bottom of the well to the surface vibrating screen. When the vision system alarms on the surface, severe stuck drill biting has often already occurred at the bottom of the well. The lack of a time delay phase difference compensation mechanism that incorporates bottom-level mechanical torque fluctuations prevents direct coordination with electrical control terminals such as the drilling rig's PLC (Programmable Logic Controller). Summary of the Invention

[0007] The purpose of this invention is to overcome the aforementioned deficiencies of the prior art and provide a closed-loop control method, device, and storage medium for wellbore anti-sticking. This invention achieves a complete reconstruction across multiple dimensions, including deep coupling of hardware and software, spatial horizon decoupling, large model attention mechanism modification, mathematical artifact filtering, and spatiotemporal delay compensation. This is followed by direct physical closed-loop control of the drilling rig at the bottom level via electrical communication.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A multimodal semantic-based anti-jamming control method, which is jointly executed by an edge drop acquisition platform, an airflow shield, a data gateway, a multimodal processing engine with a cross-feature fusion layer, and the drilling rig's electrical control terminal, includes the following steps:

[0010] Step S1: Acquire a continuous video stream of rock debris that is in free fall at the edge drop acquisition platform. Use the free fall trajectory of the rock debris after it leaves the vibrating screen to isolate the two-dimensional projection contour, so as to eliminate the adhesion of multiple targets. Extract the current video frame under the gravity discrete field of view.

[0011] Step S2: Access the standard data stream of well site information transmission at the drilling site through the data gateway, extract the current working condition data including well depth, expected lithology, surface torque and drilling fluid status, convert it into a dynamic prompt word matrix, and inject it into the cross feature fusion layer of the multimodal processing engine as a weight penalty factor through splicing and mapping to adjust the activation weight of specific geological features.

[0012] Step S3: Under the attention constraint of the weight penalty factor, the multimodal processing engine performs semantic segmentation on the current video frame, extracts the pixel-level mask of the effective abnormal blockage target, and calculates the two-dimensional physical shape size by combining the pre-calibrated camera pixel equivalent, and then calculates the visual blockage index in the current time window.

[0013] Step S4: Combine the annular cuttings return time to calculate the visual rockfall index with the time phase difference of the ground torque fluctuation in the well site information transmission standard data. Based on the evolutionary model with spatiotemporal delay compensation, perform a stuck drill risk assessment. Based on the assessment result, generate instructions that conform to the underlying control protocol and output electrical signals to the drilling rig programmable logic controller via the industrial bus to adjust the mud pump discharge or top drive speed for anti-stuck drill closed-loop control.

[0014] Compared with the prior art, the present invention has the following significant advantages:

[0015] 1. An anti-interference acquisition platform constrained by physical and fluid dynamics was constructed: by using the edge free fall trajectory and the anti-splash air curtain with a specific angle of 15° to 30°, the mutually adhering rock debris group was decoupled into an isolated two-dimensional projection in three-dimensional physical space, eliminating motion artifacts and occlusion interference caused by high-frequency vibration from the data source.

[0016] 2. A low-level attention constraint mechanism for multimodal geological semantics was created: It abandons the traditional approach of treating prompts merely as surface-level text input. Instead, it transforms standard data streams of well site information transmission into weighted penalty factors through matrix mapping equations, and deeply injects these factors into the cross-feature fusion layer of the multimodal processing engine (VLM). This enables the model to possess prior judgment capabilities similar to those of a "geological expert."

[0017] 3. An artifact filtering system with rigorous mathematical logic was established: by calculating the local curvature gradient and using spline interpolation, the perimeter distortion data caused by overlapping boundaries was completely eliminated, ensuring that the "morphological sharpness" characterizing shear failure has real engineering and physical significance.

[0018] 4. Achieved industrial-grade closed-loop electrical control across time and space misalignment: Introduced a time-space delay compensation mechanism to accurately calculate the phase difference of the annular return time. Visually detected anomalies are traced back to the moment of sudden mechanical torque change on the timeline, completely bridging the time gap between surface observation and actual wellbore conditions. Furthermore, the warning results are directly converted into PLC bus electrical signals driving the drilling rig's mud pump or top drive, achieving unmanned, proactive anti-sticking drill defense. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the overall architecture of a closed-loop control system for preventing stuck drill pipe based on multimodal geological semantic attention constraints, provided for an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the weight penalty factor injection mechanism of the cross-feature fusion layer of the multimodal processing engine provided in an embodiment of the present invention.

[0022] Figure 3 The diagram illustrates the timing verification principle of the spatiotemporal delay compensation evolution deduction model provided in this embodiment of the invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0025] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0026] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0027] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0028] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0029] This embodiment provides a closed-loop control system for preventing stuck drill bit operation that deeply couples bottom-level electrical control with multimodal AI algorithms. The system comprises an edge drop acquisition platform, an airflow shield, a data gateway, a multimodal processing engine with a cross-feature fusion layer, and the drilling rig's electrical control terminal (such as the drilling rig's PLC bus). The application scenario is set as follows: a horizontal well in an oil and gas field is being drilled to a depth of 4500 meters, currently encountering a hard and brittle shale formation, using high-viscosity oil-based drilling fluid.

[0030] Step S1: Construct an anti-interference acquisition platform with fluid dynamic constraints to acquire discrete gravity field-of-view images.

[0031] like Figure 1 As shown, this step involves deploying the explosion-proof image acquisition lens of the edge drop acquisition platform at the sand discharge end of the vibrating screen. To address drilling fluid splashing, an airflow shield (air knife system) is installed in front of the explosion-proof image acquisition lens. A specific angle of 15° to 30° is set between the airflow axis emitted by the air knife system and the main optical axis of the explosion-proof image acquisition lens. This hydrodynamic arrangement utilizes a high-pressure air curtain to cut off splashing mud without generating additional downward airflow pressure that interferes with the natural free fall motion of the cuttings.

[0032] After the rock fragments detach from the screen, they achieve physical spatial decoupling in the air by utilizing their free fall trajectory, thereby realizing the complete isolation of the two-dimensional projected contour and eliminating the computational redundancy caused by the adhesion of multiple targets on the screen from the physical source.

[0033] Step S2: Multi-source engineering data stream parsing and dynamic prompt word matrix generation.

[0034] like Figure 2 As shown, this step involves real-time access to the well site information transmission standard data stream via a data gateway. Current operating condition data is extracted: well depth is 4500 meters, expected lithology is shale, and drilling fluid condition is high viscosity. This discrete feature data is then converted into a high-dimensional dynamic cue word matrix P. matrix Its matrix mapping equation is constrained as follows:

[0035]

[0036] Among them, W geo Let E be the geological prior weight matrix, σ be the activation function, and E be the weight matrix. depth For well depth data, L lithology For the expected lithological data, M mud Here is the drilling fluid state data, b is the bias vector, and P is the bias vector. matrix This represents the multidimensional feature tensor generated after the activation function mapping, used to characterize the comprehensive geological and fluid features at the current well depth in the network's bottom layer.

[0037] When the fluid environment and geological strata change, this mapping equation will generate different text encoding dimensional features. This P matrix This will then be injected directly into the cross-feature fusion layer of the multimodal processing engine as a weighting penalty factor. This mechanism enables the model to be automatically activated under "shale" geology and to give higher attention weight to "extracting irregular angular masks" while suppressing other interfering responses.

[0038] Step S3: Geometric mapping anti-artifact mechanism and physical feature calculation.

[0039] This step extracts the pixel-level mask of the target identified as an aberrant drop, and then performs artifact filtering and physical morphology calculation. In the drop view, occasional target overlap can cause distortion at the mask boundaries. The system first calculates the local curvature gradient of the mask contour point set. If the local curvature gradient exceeds a preset high-frequency threshold, the boundary segment is determined to be an overlapping occlusion pseudo-boundary, and spline interpolation is used for smooth removal to obtain the effective pixel area (Area). effective Perimeter of the effective boundary effective .

[0040]

[0041] Combined with pre-calibrated camera pixel equivalent K scale The sharpness R of the abnormally dropped block is calculated using the following formula:

[0042] This method filters out perimeter artifacts caused by overlapping boundaries, ensuring that the morphological sharpness has physical authenticity to guide engineering closure. Furthermore, the system calculates the visual chipping index (CI) within the current time window, which is the ratio of the cumulative projected area of ​​all valid chips to the total projected area of ​​rock debris per unit time.

[0043] Step S4: Deduction of time and space delay compensation and direct PLC closed-loop control.

[0044] like Figure 3 As shown, this step completely solves the spatiotemporal misalignment problem between visual logging and the actual state at the bottom of the well by constructing an evolutionary deduction model formula. Let T0 be the actual moment of bottom-hole damage, T1 be the moment of sudden change in surface torque, and T2 be the moment when cuttings return to the edge drop zone and are detected visually. The system combines the annular cuttings return time ΔTlag (the ratio of well depth to annular return velocity) to calculate the time phase difference between visual block drop index detection and surface torque fluctuation:

[0045]

[0046] When the visual rock fragmentation index CI exceeds the preset first threshold at time T2 and the sharpness R increases, shear failure is determined to have occurred. If CI continues to rise and exceeds the second preset threshold, and high-frequency abnormal fluctuations in the surface torque parameter are found when tracing back to time T1 = T2 - ΔΦ on the time axis, the time-series evolution model determines that the rock fragment bed in the annulus at the bottom of the well has been severely thickened.

[0047] At this point, the system bypasses manual intervention, generates low-level control protocol commands (such as Modbus TCP / IP), and directly outputs adjustment signals to the drilling rig's electrical control terminal (PLC bus). For example, it outputs an electrical signal to the mud pump control module to increase the discharge rate and clean the wellbore, or an electrical signal to the top drive control module to reduce the rotation speed or stop drilling. This embodiment truly realizes a leap from data monitoring to low-level automated physical equipment anti-sticking control.

[0048] Conversely, if no high-frequency abnormal fluctuations in the ground torque parameters are detected when tracing back to time T1 on the timeline, the risk of actual stuck drill bit at the bottom of the well is ruled out. The current visual anomaly is determined to be a local block drop or visual interference. The system will continue to monitor and filter, and will not trigger the underlying electrical control.

[0049] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for preventing drill jamming based on multimodal semantics, characterized in that, The method is executed collaboratively by an edge drop acquisition platform, an airflow shield, a data gateway, a multimodal processing engine with a cross-feature fusion layer, and the drilling rig's electrical control terminal, and includes the following steps: Step S1: The explosion-proof image acquisition lens configured on the edge drop acquisition platform acquires a continuous video stream of rock debris in free fall. The free fall trajectory of the rock debris after leaving the vibrating screen is used to isolate the two-dimensional projection contour to eliminate the adhesion of multiple targets. The current video frame under the gravity discrete field of view is obtained by frame extraction. Step S2: Access the well site information transmission standard data stream at the drilling site through the data gateway, extract the current working condition data including well depth, expected lithology, surface torque and drilling fluid status, convert it into a dynamic prompt word matrix through matrix mapping equation and use it as a weight penalty factor, and inject it into the cross feature fusion layer of the multimodal processing engine. Step S3: Under the attention constraint of the weight penalty factor, the multimodal processing engine performs semantic segmentation on the current video frame, extracts the pixel-level mask of the effective abnormal falling block target, and calculates the two-dimensional physical shape size in combination with the pre-calibrated camera pixel equivalent, and then calculates the visual falling block index CI in the current time window. The visual falling block index CI is the ratio of the cumulative projected area of ​​all effective abnormal falling blocks to the total projected area of ​​rock debris per unit time. Step S4: Calculate the time phase difference between the visual rock cuttings return time and the ground torque fluctuation in the standard data stream of the well site information transmission, based on the annular cuttings return time. Perform a stuck drill risk assessment based on the evolutionary model with spatiotemporal delay compensation, and output an adjustment electrical signal directly to the drilling rig electrical control terminal according to the assessment result to perform anti-stuck drill closed-loop control.

2. The method according to claim 1, characterized in that, In step S1, the explosion-proof image acquisition lens is equipped with an airflow shield. The axis of the airflow ejected from the airflow shield forms an angle of 15° to 30° with the main optical axis of the explosion-proof image acquisition lens. This is used to maintain the hydrodynamic stability of the rock debris's natural free fall motion while physically isolating mud splashes.

3. The method according to claim 1, characterized in that, In step S2, the current operating condition data is converted into a dynamic prompt word matrix P. matrix The matrix mapping equation is: Among them, P matrix W represents the multidimensional feature cue word matrix generated after transformation. geo E is the geological prior weight matrix (offline weights obtained based on historical geological samples). depth For well depth data, L lithology For the expected lithological data, M mud Here, b represents the drilling fluid state data, σ represents the bias vector, and σ represents the activation function. The dynamic prompt word matrix P matrix At different geological strata, the cross-attention weight parameters are input into the cross-feature fusion layer to enhance the feature response of the multimodal processing engine to abnormal block drop targets that match the current working condition data, suppress interference features, and thereby extract the pixel-level mask of the abnormal block drop targets.

4. The method according to claim 1, characterized in that, In step S3, the steps of extracting the pixel-level mask of the valid abnormal block drop target and calculating the two-dimensional physical shape size specifically include: Calculate the local curvature gradient of the preliminarily extracted abnormal block mask contour point set. If the local curvature gradient exceeds the preset high frequency threshold, the corresponding contour boundary segment is determined to be an overlapping occlusion pseudo boundary. After identifying the false boundaries, the adjacent valid boundary endpoints are extracted as control points. Spline interpolation is then used to smoothly remove the overlapping occlusion false boundaries, completing the contour closure reconstruction and obtaining the effective pixel area Area. effective Perimeter of the effective boundary effective ; Combined with pre-calibrated camera pixel equivalent K scale The morphological sharpness R after overlay artifact filtering is calculated using the following formula: Among them, K scale Pre-calibration is required.

5. The method according to claim 1, characterized in that, In step S4, the process of determining the stuck drill risk using the evolutionary model with spatiotemporal delay compensation includes: Combined with the annular cuttings uplift time ΔT lag The time phase difference ΔΦ between the visual block index detection time T2 and the ground torque change time T1 is calculated. The calculation formula is ΔΦ = T2 - T1. If the visual rockfall index exceeds the preset threshold at time T2, and high-frequency abnormal fluctuations in the surface torque parameter are detected when the time axis is traced back to T1 = T2 - ΔΦ, then it is determined that the rock cuttings bed in the annulus at the bottom of the well has become abnormally thickened. The direct output of adjustment signals to the electrical control terminal of the drilling rig specifically involves: generating a low-level control protocol instruction and outputting a mud pump displacement increase signal or a top drive speed control signal to the drilling rig programmable logic controller via a bus.

6. A multimodal semantic-based anti-jamming drill control device, characterized in that, include: The fall field of view capture module is used to acquire a continuous video stream of rock debris in free fall at the edge fall acquisition platform. It uses the free fall trajectory to isolate the two-dimensional projection contour and extracts the current video frame. The matrix mapping and penalty injection module is used to extract well site information transmission standard data stream containing working condition data through the data gateway, convert it into a dynamic prompt word matrix and inject it as a weight penalty factor into the cross feature fusion layer; The feature fusion and anti-spoofing mapping module is used to perform semantic segmentation on the current video frame under the constraint of the weight penalty factor to extract the effective pixel-level mask of the abnormal block target, filter artifacts and calculate the two-dimensional physical shape size and visual block index. The phase compensation and bus control module is used to calculate the time phase difference by combining the annular return time, determine the risk of stuck drill bit based on the evolutionary model, and output adjustment electrical signals to the drilling rig electrical control terminal for physical closed-loop control.

7. An electronic device, characterized in that, include: Processor and memory; The memory stores a computer program, which, when executed by the processor, implements the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images