Wellbore facility defect cooperative detection method and device using inspection instrument lifting process

Abstract

The application discloses a wellbore facility defect cooperative detection method and device using a patrol instrument lifting process, comprising synchronously collecting visual image data and physical sensing data of the patrol instrument; running a time-varying dynamics state observer of a variable-length suspension system based on the physical sensing data to predict instantaneous vibration characteristics at a current well depth; taking the instantaneous vibration characteristics as physical priori to repair the image; and performing defect detection on the repaired image using a well depth adaptive detection mechanism. The application establishes a "well depth-dynamics-vision" cross-domain cooperative model, applies physical dynamics priori to visual processing, effectively decouples motion distortion and real defects, improves the precision of deep well defect detection, realizes the consistency of detection performance in the whole wellbore range, and solves the problem of fluctuation of detection performance with well depth.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and computer vision technology, specifically to a collaborative detection method and device for well shaft facility defects using the lifting and lowering process of an inspection instrument. Background Technology

[0002] In the daily operation and safety monitoring of deep coal mine shafts, ensuring the structural integrity of internal facilities such as shaft passages, shaft walls, and various pipelines is fundamental to ensuring safe mine production. Traditional shaft inspection methods mainly rely on workers riding in cages and conducting inspections visually or using handheld devices. These methods are not only inefficient and prone to missed inspections, but also pose potential threats to worker safety due to the harsh conditions typically present in underground environments such as high humidity, water spray, high dust concentrations, and insufficient lighting. Currently, industry technology is gradually shifting towards intelligent inspection instruments equipped with multiple sensors, including high-resolution industrial cameras, LiDAR, and thermal imagers. These instruments are raised or lowered via wire rope systems to achieve automated scanning of the entire shaft. This non-contact inspection technology significantly improves inspection efficiency and enables comprehensive digitization and visualization of shaft facilities.

[0003] However, as global mineral resource extraction extends into deeper strata, deep wells exceeding 1,000 meters in depth are becoming increasingly common. Within this vast range of depth variations, the flexible steel wire rope used for suspending inspection instruments exhibits complex nonlinear physical characteristics. Specifically, this steel wire rope system is a flexible continuum whose mass, stiffness, and length all vary with time. Its lateral stiffness, longitudinal tensile stiffness, system damping coefficient, and the natural frequency of the entire suspension system all exhibit significant nonlinear time-varying characteristics as the rope length changes. Especially during high-speed lifting, for example, when the lifting speed exceeds 5 meters per second, the complex piston wind effect within the wellbore, the periodic excitation caused by unevenness at the guide rail joints, and the mechanical vibration of the lifting equipment itself all act on this variable-length flexible suspension system. These factors collectively lead to complex six-degree-of-freedom attitude disturbances and multimodal coupled flutter in the inspection instrument. Even if the inspection instrument is equipped with a passive mechanical damping device or an active gyro-stabilized gimbal, it is difficult to completely eliminate the low-frequency large-amplitude sway (i.e., pendulum effect) determined by the dynamic characteristics of the large-scale system and the high-frequency micro-vibration related to the well depth.

[0004] Existing well inspection image processing technologies mostly focus on general denoising of static images or deblurring based on simple uniform motion models. For example, some deblurring algorithms, such as Wiener filtering, the Richardson-Lucy algorithm, or blind deblurring methods based on convolutional neural networks, typically assume that the motion blur kernel is spatially uniform and temporally invariant across the entire image or within a short timeframe, or rely solely on edge information within the image content for blind deconvolution estimation. However, in the specific application scenario of deep well inspection, the technical challenge lies in the strong nonlinear coupling between the well depth position of the inspection instrument and its exhibited dynamic response characteristics during the lifting and lowering process of a variable-length flexible suspension system. This means that the same type of well facility defect will show different results in data collected from shallow wells (where the rope is short, system stiffness is high, and natural frequency is high, with interference mainly manifested as high-frequency jitter) and deep wells (where the rope is long, system stiffness is low, and natural frequency is low, with interference mainly manifested as large-amplitude low-frequency oscillations superimposed with mid-frequency vibrations). This distortion specifically includes anisotropic motion blur at specific frequencies, non-rigid geometric distortion, and artifacts caused by the rolling shutter effect.

[0005] Existing static image processing or general denoising algorithms cannot dynamically perceive and decouple the physical motion characteristics of such time-varying systems based on the key physical parameter of "well depth." These algorithms tend to treat structured vibration disturbances caused by rope length variations as irregular random noise, resulting in significant instability in their denoising or repair effects with varying well depth. For example, in shallow wells, the algorithm may lose minute crack details due to over-smoothing; while in deep wells, it may fail to effectively remove blurring caused by large-amplitude movements. This inconsistency ultimately leads to a huge difference in defect detection performance between deep and shallow regions, with the false alarm rate or missed detection rate fluctuating nonlinearly with well depth, failing to meet the stringent requirements of coal mine safety regulations for high-precision, blind-spot-free detection throughout the entire wellbore. Therefore, there is an urgent need in the current technological field for a cross-domain coupled detection method that can synergistically consider well depth, system nonlinear dynamics, and computer vision imaging features. Summary of the Invention

[0006] One objective of this application is to provide a collaborative detection method for wellbore facility defects using the raising and lowering process of an inspection instrument, in order to solve the problem in the prior art where the image distortion features change nonlinearly with well depth due to the strong coupling between the well depth position of the inspection instrument and its dynamic response characteristics, resulting in poor consistency in defect detection between deep and shallow regions by existing static vision algorithms.

[0007] The first aspect of this application provides a method for collaborative detection of defects in well shaft facilities using the raising and lowering process of an inspection instrument, the method comprising the following steps:

[0008] 1) Synchronously acquire visual image data from the inspection instrument and physical sensor data containing current well depth information and real-time hoisting speed, and perform spatiotemporal registration on the visual image data and the physical sensor data;

[0009] 2) Based on the physical sensing data, run the time-varying dynamic state observer of the variable length suspension system and output the instantaneous vibration feature vector in real time, which includes anisotropic motion fuzzy kernel parameters, system main vibration mode frequencies, and the maximum predicted offset angle of the camera optical axis.

[0010] 3) Input the instantaneous vibration feature vector and the current well depth information as conditions into the image restoration network to restore the visual image data and output the restored image;

[0011] 4) On the repaired image, a threshold dynamic adjustment function that decreases nonlinearly with well depth is used to screen candidate defect regions to obtain determined defect regions;

[0012] 5) Combining the current well depth information and camera calibration parameters, the two-dimensional pixel coordinates of the defect area in the image are mapped to physical coordinates in the three-dimensional wellbore space coordinate system.

[0013] Furthermore, the physical sensing data also includes inertial measurement unit data; the spatiotemporal registration in step 1) specifically involves: using a hardware-triggered mode, at the center moment of each camera exposure, locking the current well depth information, real-time lifting speed, and the inertial measurement unit data, thereby mapping and binding each frame of visual image data to its corresponding physical sensing data one by one.

[0014] Furthermore, the method also includes an online calibration step for dynamic parameters based on visual residual feedback, specifically: after outputting the restored image in step 3), the physical consistency residual between the restored image after re-blurring the image using the anisotropic motion blur kernel parameters predicted in step 2) and the original input visual image data is calculated. When the physical consistency residual When the preset threshold is exceeded, the physical parameters inside the time-varying dynamic state observer are updated online using the gradient descent method with the residual value.

[0015] Preferably, the time-varying dynamic state observer in step 2) is a surrogate model constructed using a recurrent neural network based on a long short-term memory network or a gated recurrent unit to numerically solve the second-order nonlinear differential equations describing the motion of the inspection instrument.

[0016] Preferably, the image restoration network adjusts the physical consistency of the restoration process in the following manner: the image restoration network includes a multilayer perceptron that maps the instantaneous vibration feature vector to a physical feature embedding vector, a spatial feature transformation layer that modulates the visual features extracted from the visual image data using the physical feature embedding vector, and a well depth-dependent dynamic convolutional layer; wherein, the convolutional kernel weights of the dynamic convolutional layer are dynamically generated by a supernetwork, and the supernetwork takes the current well depth information and the instantaneous vibration feature vector as input.

[0017] Preferably, the threshold dynamic adjustment function The expression is: ,in, Given the current well depth, As the baseline threshold, The maximum depth of the wellbore. These are hyperparameters used for decay rate control. This is the minimum safety threshold.

[0018] Preferably, the image restoration network is trained using a physical consistency closed-loop training strategy based on synthetic data. This strategy includes: generating training data containing well depth, images with physical distortion characteristics, and distortion-free clear images using a computer graphics engine; during training, a joint loss function is employed, which includes a physical consistency re-blur loss. This re-blur loss is calculated by performing a re-convolution blur operation on the restored image output by the network using the predicted anisotropic motion blur kernel parameters to obtain a re-blurred image, and calculating the pixel-level difference between the re-blurred image and the input image with physical distortion characteristics.

[0019] A second aspect of this application provides a collaborative detection device for defects in well shaft facilities utilizing the lifting and lowering process of an inspection instrument, the device comprising:

[0020] The multi-source cross-domain synchronous acquisition module is used to synchronously acquire visual image data from the inspection instrument and physical sensor data containing current well depth information and real-time lifting speed, and perform spatiotemporal registration.

[0021] The time-varying dynamic state observation module is used to run the time-varying dynamic state observer of the variable length suspension system based on the physical sensing data, and output the instantaneous vibration feature vector containing anisotropic motion fuzzy kernel parameters in real time.

[0022] The physical-guided image restoration module has a built-in image restoration network, which is used to input the instantaneous vibration feature vector and the current well depth information as conditions into the image restoration network, restore the visual image data and output the restored image;

[0023] The well depth sensing collaborative detection module is used to screen candidate defect regions on the repaired image using a threshold dynamic adjustment function that decreases nonlinearly with well depth, to obtain determined defect regions, and to map the two-dimensional pixel coordinates of the defect regions in the image to physical coordinates in the three-dimensional wellbore space coordinate system by combining the current well depth information and camera calibration parameters.

[0024] Furthermore, the image restoration network in the physically guided image restoration module adjusts the physical consistency of the restoration process in the following way: the image restoration network includes a multilayer perceptron that maps the instantaneous vibration feature vector to a physical feature embedding vector, a spatial feature transformation layer that uses the physical feature embedding vector to perform affine transformation modulation on the visual features extracted from the visual image data, and a well depth-dependent dynamic convolutional layer; wherein, the convolutional kernel weights of the dynamic convolutional layer are dynamically generated by a supernetwork, and the supernetwork takes the current well depth information and the instantaneous vibration feature vector as input.

[0025] The beneficial effects of this application are as follows:

[0026] 1. This application establishes a cross-domain collaborative model of "well depth-dynamics-vision", applies the prior knowledge of physical dynamics to visual processing, and uses a time-varying dynamic state observer to deduce features such as motion fuzzy kernel in real time, effectively decoupling the motion distortion introduced by the variable length suspension system from the actual defect features of the facility;

[0027] 2. This application proposes an online calibration mechanism for dynamic parameters based on visual residual feedback, which can update the dynamic model parameters online according to the physical consistency residual of image restoration, thereby enhancing the system's robustness to equipment aging and environmental changes;

[0028] 3. This application employs a depth-dependent dynamic convolutional layer and a threshold dynamic adjustment function that decreases nonlinearly with depth, thereby achieving adaptive processing of different imaging features in deep and shallow areas. This ensures the consistency of defect detection performance across the entire wellbore and effectively reduces the false alarm rate in deep areas and the false alarm rate in shallow areas.

[0029] 4. This application constructs a closed-loop training strategy based on physical consistency reprojection loss, which solves the problem of scarce real deep well defect samples by using synthetic data generated by a virtual engine, and forces the network to learn the inverse process of physical degradation, thereby improving the physical authenticity of the repair. Attached Figure Description

[0030] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0031] Figure 1 This is a flowchart illustrating a collaborative detection method for wellbore facility defects using the lifting and lowering process of an inspection instrument, as provided in an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of a scenario in which the inspection instrument performs multi-source cross-domain synchronous data acquisition inside the well shaft, according to an embodiment of the present invention.

[0033] Figure 3 This is the core logic block diagram of the "well depth-dynamics-vision" cross-domain collaborative processing in the embodiments of the present invention.

[0034] Figure 4 This is a schematic diagram of the structure of the multi-scale defect collaborative detection module based on well depth sensing threshold in an embodiment of the present invention.

[0035] Figure 5 This is a characteristic curve of the lateral stiffness of the suspension system and the natural frequency of the system as a function of well depth in an embodiment of the present invention.

[0036] Figure 6 This is a schematic diagram comparing the motion fuzzy kernel morphology output by the time-varying dynamic state observer at different well depths in an embodiment of the present invention, wherein sub-figure (a) is the shallow case; sub-figure (b) is the middle case; and sub-figure (c) is the deep case.

[0037] Figure 7 This is a comparison chart of the defect detection performance of the method of the present invention and the prior art in different well depth sections in the embodiments of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0039] Example 1:

[0040] This embodiment provides a collaborative detection method for wellbore facility defects utilizing the raising and lowering process of an inspection instrument, referring to... Figure 1 The process shown is designed to address the inconsistency in visual inspection performance caused by variations in the dynamic characteristics of the inspection instrument at different well depths.

[0041] Step one: Perform multi-source, cross-domain synchronous data acquisition and unify the spatiotemporal reference. (Refer to...) Figure 2 As the inspection instrument moves up and down the shaft at full depth, the system simultaneously initiates data acquisition. Specifically, using a high-speed industrial camera array arranged circumferentially around the inspection instrument body, it continuously acquires video stream sequences of facilities such as the inner wall of the shaft, tank passages, and pipelines at a preset frame rate, forming the original visual input matrix. Meanwhile, the encoder, linked to the wire rope drum of the hoisting system, collects and outputs the current well depth information of the inspection instrument in real time. and lifting speed In addition, to capture the attitude disturbances of the inspection instrument body, a high-precision six-axis inertial measurement unit (IMU) fixed to the camera module base acquires linear acceleration and angular velocity data along three axes at a frequency higher than the camera frame rate (e.g., 1000 Hz), forming an inertial measurement data stream. To ensure accurate temporal correspondence between physical sensing data and visual data, this embodiment employs a hardware-triggered synchronization mechanism. Specifically, the center moment of the camera exposure pulse is used as the reference time point, and the encoder reading and real-time speed at that moment are immediately latched by the hardware circuit. And IMU data samples. In this way, each frame of the acquired image... All correspond precisely to the physical state vector By performing one-to-one mapping and binding, structured cross-domain data pairs are constructed, providing a data foundation for subsequent physical-visual collaborative analysis.

[0042] Step 2: Build and run a time-varying dynamic state observer for the variable-length suspension system. (Refer to...) Figure 3 The core of this step lies in using the real-time acquired physical state vectors. This allows for real-time deduction of the instantaneous vibration characteristics of the downhole inspection instrument at the current well depth, thus providing a physical prior for subsequent image restoration. Specifically, a dynamic model of the suspension system is first established. Based on the continuum mechanics theory of variable-length flexible cables, the suspension wire rope is equivalent to a beam model whose length varies with time, and its equivalent lateral stiffness... and longitudinal stiffness All are well depths The function is related to the tension at depth caused by the weight of the inspection device and the wire rope, the elastic modulus of the wire rope, and the moment of inertia of the wire rope cross section. Based on this, a set of second-order nonlinear differential equations describing the six-degree-of-freedom motion of the inspection device is established. This set of equations includes the generalized mass matrix, displacement and rotation vectors, time-varying damping matrix (including air damping, etc.), time-varying stiffness matrix, and external excitation forces (such as piston wind effect or guide rail joint impact). Since this set of equations is difficult to solve analytically in real time, this embodiment uses a recurrent neural network based on gated recurrent units (GRU) as a surrogate model for this dynamic system.

[0043] like Figure 5 As shown, this graph illustrates the lateral stiffness of the suspension system. and the system's natural frequency As well depth The relationship is nonlinear and varies. The horizontal axis represents the well depth. The range is from 0 to 1200 meters; the left vertical axis represents the lateral stiffness. The unit is N / m; the right vertical axis represents the system's natural frequency. The unit is Hz. The solid curve represents the variation of lateral stiffness with well depth: In the shallow region (0 to 200 meters), the lateral stiffness remains at a high level and changes gradually, approximately 8000 to 8500 N / m; in the transition region (200 to 800 meters), the lateral stiffness decreases rapidly and nonlinearly with increasing well depth, indicating that the lateral restraint capacity of the suspension system is significantly weakened in this range; in the deep region (800 to 1200 meters), the lateral stiffness tends to be low and flat, approximately 800 to 1200 N / m, and the lateral support capacity of the system has been greatly reduced. The dashed curve represents the system's natural frequency. The overall attenuation pattern with well depth is similar to that of lateral stiffness, but the attenuation rate is faster, with the natural frequency rapidly decreasing from approximately 5.2 Hz in shallow wells to approximately 0.35 Hz in deeper wells. These characteristics indicate that as well depth increases, the lateral stability and vibration resistance of the suspension system gradually decrease, necessitating targeted vibration suppression measures to ensure safe system operation in deep well conditions.

[0044] Normalized well depth ,speed and IMU data As input, the pre-trained GRU network can output the instantaneous vibration feature vector of the inspection instrument at the current moment in real time. Preferably, the vector explicitly includes three key physical quantities: the system's principal vibrational mode frequencies. It characterizes the frequency of periodic ripples or distortions that may appear in the image; the maximum predicted offset angle of the camera optical axis. It quantifies the overall geometric distortion amplitude caused by low-frequency oscillations; and the anisotropic motion fuzzy kernel parameters. This parameter is determined by the fuzzy length. (Unit: pixels) and blur angle (Unit: degrees) constitutes a precise description of the linear blur caused by camera movement within the current frame's exposure time.

[0045] like Figure 6 As shown, the morphology of motion blur kernels at different well depths is compared. (a) shows that at shallow depths (h=180m), the blur kernels exhibit a small and dense dotted distribution, indicating that the influence is mainly high-frequency micro-vibrations, resulting in a small blur scale (L≈5 pixels); (b) shows that at middle depths (h=600m), the blur kernels exhibit an elliptical distribution, indicating the influence of mid-frequency vibrations, resulting in an increased blur scale (L≈12 pixels); (c) shows that at deep depths (h=1150m), the blur kernels exhibit a significant elongated linear feature, indicating that the influence is dominated by low-frequency large-amplitude oscillations, resulting in the largest blur scale (L≈18 pixels) and a specific directionality. (degree). This figure visually reflects the nonlinear influence of well depth variation on imaging quality.

[0046] Furthermore, to address the problem of physical parameter mismatch in actual working conditions, this embodiment introduces an online calibration mechanism for dynamic parameters based on visual residual feedback. For example... Figure 3 As shown in the feedback loop, when the subsequent image restoration module (step three) generates the restored image... Then, the system will calculate the physical consistency residual. The residual is calculated as follows: first, using the fuzzy kernel predicted by the current dynamic state observer... For image restoration Perform convolution operations to simulate the physical blurring process and obtain a reblurred image. Then calculate the reblurred image. Compared with the original input image The L1 norm or L2 norm distance between them, i.e. If the residual If the value remains consistently higher than a preset threshold, it indicates a discrepancy between the observer's predicted physical process and the actual situation. In this case, the system will utilize this residual value. The gradient descent method is used to calculate the gradients of key physical parameters in the dynamic differential equations and to fine-tune and update them. This closed-loop calibration mechanism constructs a feedback loop, using the residuals from image inpainting to calibrate the physical parameters of the dynamic model, enabling the dynamic model to self-correct based on visual results.

[0047] Step 3: Perform well depth adaptive image inpainting based on physical prior feature embedding. The goal of this step is to utilize the instantaneous vibration feature vector output from Step 2. As powerful prior knowledge, it is relevant to the original perturbed image. Physically consistent repairs are performed. This embodiment employs an improved physics-guided generative adversarial network (GAN) to achieve this functionality, which includes a generator and a discriminator. Specifically, the instantaneous vibration feature vector is first... Inputting the data into a multilayer perceptron (MLP) network maps it to a high-dimensional physical feature embedding vector. Subsequently, Spatial Feature Transform (SFT) layers are introduced into multiple layers of the encoder and decoder of the generator network. The SFT layers utilize physically embedded vectors. A pair of scale and bias parameters is generated, and the visual feature map of this layer is modulated element-wise by an affine transformation. This modulation mechanism enables each layer of the generator to perceive the current well depth, vibration frequency, and blur state, thereby performing targeted feature inpainting. Furthermore, to handle distortion patterns strongly correlated with well depth more precisely, a well depth-dependent dynamic convolutional layer is introduced in the key recovery layer of the generator network's decoder. The convolutional kernel weights of this layer are not fixed and pre-trained, but are dynamically generated in real-time by a lightweight hypernetwork that is based on well depth. and vibration characteristics For input.

[0048] For example, when the inspection instrument is located deep within the well shaft ( When the value is large, low-frequency large oscillations are the main interference. At this time, the supernetwork will generate convolution kernels with large receptive fields and anisotropic shapes to effectively correct large geometric distortions; while when the inspection instrument is in a shallow area ( When the value is small, high-frequency micro-vibrations are the main interference. The supernetwork generates sharpening convolution kernels similar to the Laplacian operator to accurately eliminate high-frequency motion blur. Through the above mechanism, the generator finally outputs a high-quality restored image that is decoupled from physical dynamics. Its blurring, noise and geometric distortion are all effectively suppressed.

[0049] Step four: Perform multi-scale collaborative defect detection based on well depth sensing thresholds. (Refer to...) Figure 4 High-quality images after physical consistency restoration The system performs minor defect detection on the wellbore facilities. To address the inherent signal-to-noise ratio difference between the deep and shallow sections of the wellbore due to physical factors such as light attenuation and dust concentration, this embodiment introduces a depth-adaptive detection strategy. First, the repaired image... The input is fed into the backbone of a convolutional neural network-based defect detection network to extract image features. Then, a multi-scale feature map is constructed using a feature pyramid network (FPN). At the output of the detection network, the core design element of this invention lies in a threshold dynamic adjustment function that nonlinearly and monotonically decreases with well depth. A specific form of this function could be: ,in It is the baseline threshold. It is the maximum depth of the wellbore. This is the current well depth. It is a hyperparameter that controls the decay rate. This is the minimum safe threshold to prevent the threshold from being too low. The function specifies that in deep regions with low signal quality, the system automatically lowers the detection threshold to increase sensitivity to defect signals; while in shallow regions with good signal quality, a higher threshold is maintained. Finally, this is combined with the precise well depth corresponding to the frame image. Speed ​​increase By combining the camera calibration parameters, the two-dimensional pixel coordinates in the image are converted into physical coordinates in the three-dimensional spatial coordinate system of the wellbore, and finally a three-dimensional wellbore defect distribution map containing defect type, location and size information is generated.

[0050] Step 5: Perform closed-loop training based on synthetic data to ensure physical consistency. To effectively train the deep learning network, this embodiment constructs a closed-loop training strategy. Specifically, a virtual wellbore 3D environment is constructed using a computer graphics engine and physical simulation components. The real dynamic motion trajectory of the inspection instrument at different depths is simulated. The virtual camera generates image sequences with physical distortion features as input data for the network. Simultaneously, the engine outputs a clear image without motion distortion at the same location as the ground truth for supervision. And the corresponding well depth, velocity, IMU and other physical motion parameters. During network training, a joint loss function is used for optimization, which includes adversarial loss, perceptual loss, and physically consistent re-blurring loss. The core function is the physically consistent re-blurring loss. The calculation method is as follows: the repaired image output by the generator... Using the corresponding fuzzy kernel predicted in step two Perform convolution blur again to obtain a re-blurred image. and calculate With input distorted image The pixel-level differences between them. This loss term forces the generator to learn the exact inverse process of the physical degradation process, ensuring the physical authenticity of the repair results.

[0051] Example 2:

[0052] This embodiment provides a collaborative detection device for wellbore facility defects utilizing the raising and lowering process of an inspection instrument. The device's structural design is intended to perform the method described in Embodiment 1. Specifically, the device includes:

[0053] The multi-source cross-domain synchronous acquisition module integrates a high-speed industrial camera array, an encoder linked to the lifting system, a high-precision six-axis inertial measurement unit (IMU), and a hardware synchronization triggering unit. This module is responsible for controlling the industrial camera array to acquire visual images during the inspection instrument's ascent and descent, and simultaneously acquiring well depth, velocity, and attitude data through the encoder and IMU. The hardware synchronization triggering unit ensures that each frame of image is precisely spatiotemporally aligned with the physical state data at that moment, down to the microsecond level.

[0054] The time-varying dynamic state observation module typically consists of a high-performance embedded processor and memory. Its core function is to receive real-time well depth and motion data transmitted from the acquisition module and run a pre-trained time-varying dynamic state observer for the variable-length suspension system (such as a GRU-based surrogate model) within it. This module can deduce and output instantaneous vibration modal parameters and fuzzy kernel features at the current well depth in real time, and execute online dynamic parameter calibration logic based on visual residual feedback.

[0055] The physics-guided image restoration module, typically implemented using a dedicated graphics processing unit (GPU) or neural network processing unit (NPU), meets the demands of large-scale parallel computing. This module incorporates the physics-guided generative adversarial network described in Example 1. It receives the original perturbed image and the physical feature vector output by the dynamics observation module, and utilizes these physical features as prior conditions. Through well-depth attention mechanisms (such as SFT layers) and dynamic convolutional layers, it performs efficient and adaptive deblurring and geometric correction on the perturbed image, outputting a high-quality restored image.

[0056] The well depth sensing collaborative detection module, consisting of one or more processors, is responsible for executing detection tasks. It receives high-quality images output by the repair module and uses a built-in multi-scale defect detection network (such as an FPN-based detector) for defect identification and localization. The core function of this module is to retrieve an adaptive confidence threshold corresponding to the current depth from a preset threshold dynamic adjustment function based on real-time well depth information provided by the acquisition module, thereby filtering the detection results and achieving consistent and reliable detection performance across the entire well depth range. Finally, this module is also responsible for mapping the detected two-dimensional defect locations to a three-dimensional wellbore spatial coordinate system.

[0057] The above modules can be implemented through hardware circuits, software programs, or a combination thereof. They work together to form a complete intelligent detection system capable of dynamically adjusting the processing strategy according to the well depth.

[0058] Example 3:

[0059] This embodiment further illustrates the implementation process and technical effects of the method of the present invention through a specific application scenario. The goal of this application scenario is to conduct structural health monitoring of a vertical main shaft with a depth of 1200 meters, focusing on detecting whether there are fatigue microcracks in the fixing bolts at the connection between the guide beam and the shaft wall.

[0060] Due to the great depth of the wellbore and the requirement for the inspection instrument to ascend and descend at a speed of no less than 6 meters per second to ensure detection efficiency, the shaking and vibration characteristics of the inspection instrument vary greatly at different depths. When testing with existing technologies based on a general deblurring algorithm and a fixed threshold detector, it was found that at deep depths (e.g., below 1000 meters), large-amplitude low-frequency oscillations caused severe non-uniform motion blur in the image, making it impossible for conventional methods to effectively restore the image, resulting in the complete missed detection of microcracks. Conversely, at shallow depths (e.g., above 200 meters), high-frequency flutter was incorrectly processed by conventional methods, and the residual image artifacts were frequently misidentified as cracks by the detector, generating numerous false alarms. This indicates that conventional techniques are unable to adapt to the drastically changing image degradation patterns with varying well depths in this scenario.

[0061] To address the aforementioned problems, the method of this invention is applied to this scenario. First, during the descent of the inspection instrument at a speed of 6 meters per second, the multi-source cross-domain synchronous acquisition module, as described in step one, synchronously records the image sequence, well depth, velocity, and IMU data of the entire wellbore.

[0062] When the inspection instrument reaches a depth of 1150 meters, the time-varying dynamic state observation module receives the physical state vector at that depth. Based on its internal dynamic model, the module deduces in real time that the dominant disturbance of the system at this depth is a low-frequency pendulum swing with a period of approximately 2 seconds, and generates an instantaneous vibration feature vector accordingly. This vector clearly indicates that the motion blur kernel corresponding to the current frame image is a linear blur kernel with a length of 18 pixels and an orientation angle of 8 degrees, and predicts a maximum angular offset of approximately 1.5 degrees in the camera's optical axis. This physical feature vector is then fed into the physically guided image restoration module. The dynamic convolutional layer in the restoration module dynamically generates a convolutional kernel with a large receptive field and directionality based on this input, performing non-rigid geometric correction and directional deblurring on the original image.

[0063] Subsequently, when the inspection instrument ascended to a depth of 180 meters, the dynamic state observation module, based on the new physical state vector, deduced that the dominant disturbance of the system had transformed into a high-frequency micro-vibration with a frequency of approximately 9 Hz. Its output instantaneous vibration feature vector described a point spread function that was smaller in size but more complex in shape. Upon receiving this feature, the physical-guided image inpainting module correspondingly generated a high-pass filtered convolutional kernel with sharpening properties in its dynamic convolutional layer, effectively suppressing the blurring introduced by the high-frequency jitter.

[0064] After the above processing, images acquired from both deep and shallow wells are restored to clear, high-quality images with consistent geometry. Finally, the well depth sensing collaborative detection module performs defect detection on these restored images. When processing images at a depth of 1150 meters, due to the large current well depth, the threshold dynamic adjustment function... The detection confidence threshold was automatically adjusted from the baseline of 0.85 to 0.72, allowing fatigue crack candidate boxes with a width of only 0.6 mm to be retained and validated. When processing images at a depth of 180 meters, the detection threshold remained at a higher value of 0.85, effectively avoiding false positives of residual noise that might exist in the repaired image as a defect.

[0065] By implementing the method of this invention, a high-precision three-dimensional defect distribution map covering the entire wellbore is finally generated, accurately pinpointing the location of all previously missed or falsely reported microcracks in the fixing bolts, and realizing consistent and reliable defect detection across the entire well depth under high-speed inspection conditions.

[0066] like Figure 7 As shown, a comparative experiment was conducted to examine the defect detection performance of the method of this invention and existing technologies in different well depth sections. The wellbore was divided into four sections according to depth: 0-300m shallow section, 300-600m mid-shallow section, 600-900m mid-deep section, and 900-1200m deep section. Regarding detection accuracy, existing technologies, employing a fixed threshold and a general deblurring method, exhibit a significant decreasing trend in detection accuracy with increasing well depth. The accuracy is approximately 92% in the shallow section, drops to approximately 85% in the mid-shallow section, further decreases to approximately 72% in the mid-deep section, and reaches only approximately 58% in the deep section. In contrast, the method of this invention achieves detection accuracies of approximately 95%, 96%, 94%, and 93% in the four well depth sections, respectively, demonstrating consistent and uniform performance without exhibiting performance degradation with increasing depth. Regarding the false negative rate, the false negative rate of existing technologies rises sharply from about 5% in shallow wells to about 38% in deep wells, while the false negative rate of the method of this invention remains at a low level, at approximately 3%, 2.5%, 4%, and 5% in the four sections. The above comparison results demonstrate that this invention, through its adaptive parameter adjustment mechanism, effectively overcomes the detection difficulties caused by signal attenuation and noise enhancement in deep well sections, achieving consistent and reliable defect detection performance throughout the entire wellbore.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A collaborative detection method for defects in well shaft facilities utilizing the lifting and lowering process of an inspection instrument, characterized in that, Includes the following steps: 1) Synchronously acquire visual image data from the inspection instrument and physical sensor data containing current well depth information, real-time hoisting speed and inertial measurement unit data, and perform spatiotemporal registration on the visual image data and the physical sensor data; 2) Based on the physical sensing data, run the time-varying dynamic state observer of the variable length suspension system and output the instantaneous vibration feature vector in real time, which includes anisotropic motion fuzzy kernel parameters, system main vibration mode frequencies, and the maximum predicted offset angle of the camera optical axis. 3) The instantaneous vibration feature vector and the current well depth information are used as inputs into the image restoration network to restore the visual image data and output a restored image. The image restoration network adjusts the physical consistency of the restoration process in the following way: The image restoration network includes a multilayer perceptron that maps the instantaneous vibration feature vector to a physical feature embedding vector, a spatial feature transformation layer that uses the physical feature embedding vector to perform affine transformation modulation on the visual features extracted from the visual image data, and a well depth-dependent dynamic convolutional layer; wherein, the convolutional kernel weights of the dynamic convolutional layer are dynamically generated by a supernetwork, and the supernetwork takes the current well depth information and the instantaneous vibration feature vector as input. 4) After outputting the repaired image in step 3), perform online calibration of dynamic parameters based on visual residual feedback: calculate the physical consistency residual between the repaired image after re-blurring the image obtained by the anisotropic motion blur kernel parameters predicted in step 2) and the original input visual image data; when the physical consistency residual exceeds a preset threshold, use the physical consistency residual to update the physical parameters inside the time-varying dynamic state observer online using the gradient descent method; 5) On the repaired image, a threshold dynamic adjustment function that decreases nonlinearly with well depth is used to screen candidate defect regions to obtain determined defect regions; 6) Combining the current well depth information and camera calibration parameters, the two-dimensional pixel coordinates of the defect area in the image are mapped to physical coordinates in the three-dimensional wellbore space coordinate system.

2. The method according to claim 1, characterized in that, The spatiotemporal registration in step 1) specifically involves: using a hardware-triggered mode, at the center moment of each camera exposure, locking the current well depth information, real-time lifting speed, and the inertial measurement unit data, thereby mapping and binding each frame of visual image data to its corresponding physical sensor data one by one.

3. The method according to claim 1, characterized in that, The time-varying dynamic state observer in step 2) is a surrogate model constructed using a recurrent neural network based on a long short-term memory network or a gated recurrent unit to numerically solve the second-order nonlinear differential equations describing the motion of the inspection instrument.

4. The method according to claim 1, characterized in that, The expression for the threshold dynamic adjustment function is: ,in, Given the current well depth, As the baseline threshold, The maximum depth of the wellbore. These are hyperparameters used for decay rate control. This is the minimum safety threshold.

5. The method according to claim 1, characterized in that, The image restoration network is trained using a physical consistency closed-loop training strategy based on synthetic data. This strategy includes: generating training data containing well depth, images with physical distortion characteristics, and distortion-free images using a computer graphics engine; during training, a joint loss function is employed, which includes a physical consistency re-blur loss. This re-blur loss is calculated by performing a re-convolution blur operation on the restored image output by the network using the predicted anisotropic motion blur kernel parameters to obtain a re-blurred image, and then calculating the pixel-level difference between the re-blurred image and the input image with physical distortion characteristics.

6. A collaborative detection device for defects in well shaft facilities utilizing the lifting and lowering process of an inspection instrument, characterized in that, include: The multi-source cross-domain synchronous acquisition module is used to synchronously acquire visual image data from the inspection instrument and physical sensor data containing current well depth information, real-time hoisting speed and inertial measurement unit data, and perform spatiotemporal registration. The time-varying dynamic state observation module is used to run the time-varying dynamic state observer of the variable length suspension system based on the physical sensing data, and outputs an instantaneous vibration feature vector containing anisotropic motion fuzzy kernel parameters, system main vibration mode frequencies, and the maximum predicted offset angle of the camera optical axis in real time. A physically guided image restoration module, with a built-in image restoration network, is used to input the instantaneous vibration feature vector and the current well depth information as conditions into the image restoration network to restore the visual image data and output a restored image. The image restoration network adjusts the physical consistency of the restoration process in the following way: the image restoration network includes a multilayer perceptron that maps the instantaneous vibration feature vector to a physical feature embedding vector, a spatial feature transformation layer that uses the physical feature embedding vector to perform affine transformation modulation on the visual features extracted from the visual image data, and a well depth-dependent dynamic convolutional layer; wherein, the convolutional kernel weights of the dynamic convolutional layer are dynamically generated by a supernetwork, and the supernetwork takes the current well depth information and the instantaneous vibration feature vector as input; The dynamic parameter online calibration module is used to calculate the physical consistency residual between the image after the physical guided image restoration module outputs the restored image and the original input visual image data after the image is re-blurred by the anisotropic motion blur kernel parameters predicted by the time-varying dynamic state observation module. When the physical consistency residual exceeds a preset threshold, the physical parameters inside the time-varying dynamic state observer are updated online using the gradient descent method based on the physical consistency residual. The well depth sensing collaborative detection module is used to screen candidate defect regions on the repaired image using a threshold dynamic adjustment function that decreases nonlinearly with well depth, to obtain determined defect regions, and to map the two-dimensional pixel coordinates of the defect regions in the image to physical coordinates in the three-dimensional wellbore space coordinate system by combining the current well depth information and camera calibration parameters.

7. The apparatus according to claim 6, characterized in that, The expression for the threshold dynamic adjustment function is: ,in, Given the current well depth, As the baseline threshold, The maximum depth of the wellbore. These are hyperparameters used for decay rate control. This is the minimum safety threshold.

Images