Artificial intelligence-based flexible marine hydrogen pipeline crack identification method

Abstract

The application discloses a flexible marine hydrogen pipeline crack identification method based on artificial intelligence, and particularly relates to the field of machine vision identification, and is used for solving the problem of crack misjudgment caused by flow-induced vibration of the flexible pipeline; an edge computing unit carried by an underwater inspection platform is used to acquire video stream frames, spatial pose data and pipe body center line sampling points of the target flexible pipeline, an inverse kinematics compensation operator is used to deduct apparent deviation caused by the motion of the underwater inspection platform itself, a time domain physical deformation scalar is generated, a Gaussian projection kernel function, a deformation impact jump value, a frequency domain residual saliency and a support vector data description model are used to jointly generate a feature mask, the feature mask is subjected to pixel-by-pixel feature modulation with an initial two-dimensional visual feature map, a coupled feature map is formed, and after structure fracture feature extraction, a crack identification and positioning result is generated, and a link execution state is output according to the comparison result of the energy main frequency and the static water area dead zone threshold value.

Description

Technical Field

[0001] This invention relates to the field of machine vision recognition, and more specifically, to an artificial intelligence-based method for identifying cracks in flexible marine hydrogen transport pipelines. Background Technology

[0002] With the advancement of deep-sea energy projects, flexible marine hydrogen pipelines are undertaking hydrogen transportation functions under complex sea conditions. Unlike rigid onshore pipelines or fixed nearshore pipelines, flexible marine hydrogen pipelines are inherently flexible, capable of slight swaying and local expansion and contraction in response to flow fields. During long-term service, the pipeline is not only affected by internal transport pressure, external water pressure, and assembly constraints, but also continuously subjected to ocean currents, swells, inspection disturbances, and local impact loads. These factors can cause wrinkles, texture misalignment, localized tensile marks, or micro-fracture areas on the pipeline surface. For hydrogen pipelines, crack identification requires not only detecting linear fractures on the pipeline surface but also distinguishing between genuine cracks and pseudo-cracks caused by flexible deformation.

[0003] Existing pipeline defect detection typically relies on visual acquisition and image recognition. Underwater inspection platforms equipped with machine vision modules move along the outside of the pipeline, acquiring video streams containing the pipe structure. Suspected crack areas are then identified through edge detection, texture analysis, or machine learning models. This approach is usable in scenarios where the target pipeline's attitude is relatively stable, imaging conditions are clear, and the inspection platform's movement is relatively smooth. However, it implicitly assumes that texture variations in the image primarily originate from defects within the pipeline itself, rather than from the observation platform's movement or the pipe's flexible swaying. The inspection scenario for deep-sea flexible marine hydrogen pipelines does not meet this prerequisite.

[0004] In actual underwater inspections, the apparent motion of the pipe in the video stream is usually formed by the superposition of at least two types of factors: one is the movement trajectory of the underwater inspection platform itself, including thruster corrections, attitude yaw, pitch changes, and position drift; the other is the relative motion of the target flexible pipe caused by ocean currents, including pipe bending, axial tension, local oscillation, and flow-induced vibration. The video stream acquired by the machine vision module cannot naturally distinguish between these two types of motion sources. If crack identification is performed directly on the original video stream, the model will input the edge trailing caused by the underwater inspection platform's own motion, the texture ghosting caused by the pipe bending, and the local frequency domain anomalies caused by water turbidity into the crack identification link.

[0005] The problems arising from this hybrid motion are not simply image blurring. The surface of flexible marine hydrogen pipelines may contain coating textures, manufacturing seams, shadows from attachments, and pipe wrinkles. When the pipe experiences high-frequency flow-induced vibrations in ocean currents, these textures can cause nonlinear spatial misalignments in adjacent video frames. The visual model receives a high-gradient region that appears to be a "linear fracture," but this high-gradient region may simply be edge overlap or texture stretching caused by elastic deformation. Mistaking this for a crack will result in a large number of false positives; excessively smoothing the image to suppress false positives may weaken the edge response of real, fine cracks. Therefore, existing pipeline crack identification methods that only use video images as input struggle to reliably distinguish between real cracks and deformation artifacts in scenarios where the underwater inspection platform's own motion and the flexible pipe's flow-induced vibrations coexist.

[0006] Furthermore, underwater inspection platforms are typically limited by battery capacity, edge computing chip power consumption, and underwater communication bandwidth. Continuously performing complex mask calculations, frequency domain evaluations, and anomaly model analyses when the pipe is under low vibration or relatively stable conditions would result in unnecessary computational overhead. Conversely, under strong current conditions, without assessment of mask reliability, incorrectly generated masks could suppress the actual crack response. Therefore, a crack identification method is needed that can simultaneously utilize video stream frames, spatial pose data, and pipe geometric deformation information, creating a consistent processing chain between pipe physical deformation, image frequency domain state, and visual feature modulation.

[0007] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an artificial intelligence-based method for identifying cracks in flexible marine hydrogen transport pipelines. This method utilizes an edge computing unit mounted on an underwater inspection platform to acquire video stream frames, spatial pose data, and sampling points along the pipeline's centerline. An inverse kinematics compensation operator is used to subtract the apparent offset caused by the underwater inspection platform's own motion, generating a temporal physical deformation scalar. A feature mask is then generated using a Gaussian projective kernel function, deformation impact jump value, frequency domain residual significance, and a support vector data description model. The feature mask is then coupled pixel-by-pixel with an initial two-dimensional visual feature map to form a coupled feature map. After extracting structural fracture features, a crack identification and localization result is generated. Simultaneously, the link execution status is output based on a comparison between the energy dominant frequency and the static water dead zone threshold, thus addressing the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence-based method for identifying cracks in flexible marine hydrogen transport pipelines includes the following steps: Step S1: In response to the inspection command, acquire video stream frames and spatial pose data containing the target flexible pipe, extract the sampling points of the pipe centerline, and use the inverse kinematics compensation operator to compensate and generate a time-domain physical deformation scalar. Step S2: Receive the temporal physical deformation scalar and video stream frames, generate an initial two-dimensional spatiotemporal deformation probability tensor through the Gaussian projective kernel function, and perform a coupling reliability assessment based on the temporal physical deformation scalar and video stream frames to determine the feature mask; Step S3: Extract the initial two-dimensional visual feature map of the video stream frame, perform pixel-by-pixel feature modulation on the feature mask and the initial two-dimensional visual feature map in the spatial and channel dimensions, and output the coupled feature map; Step S4: Extract structural fracture features based on the coupled feature map and generate crack identification and location results, and output the link execution status based on the energy main frequency of the time-domain physical deformation scalar and the static water dead zone threshold.

[0010] Furthermore, in step S1, the edge computing unit uses the inspection command trigger time as a common time reference, writes the video stream frame, spatial pose data and pipe centerline sampling point at the same sampling time into the frame pose synchronization node, and retains the sampling time axis; the pipe centerline sampling point is formed by connecting the midpoints of the two sides of the target flexible pipe at the same axial position in axial order, and inputs the inverse kinematics compensation operator.

[0011] Furthermore, in step S1, the edge computing unit converts the attitude quaternions in the spatial pose data into rotation matrices, performs orientation verification using Euler angles, and then constructs a rigid body homogeneous transformation matrix by combining the position vector. The rigid body homogeneous transformation matrix is ​​inverted to form inverse kinematic compensation operators arranged in the order of sampling time. The inverse kinematic compensation operators are applied to the sampling points of the pipe centerline to deduct the apparent offset caused by the movement of the underwater inspection platform.

[0012] Furthermore, in step S1, the edge computing unit calculates the change in direction angle and arc length interval of adjacent centerline segments based on the sampling points of the tube centerline after deducting the apparent offset by the inverse kinematics compensation operator to obtain the macroscopic bending curvature, and calculates the change in length of the same centerline segment at adjacent sampling times to obtain the transient stretching variable, and then generates the time-domain physical deformation scalar according to the sampling time sequence.

[0013] Furthermore, in step S2, the edge computing unit uses the sampling point on the center line of the tube where the macroscopic curvature reaches its maximum value at the current sampling time as the deformation center position. Based on the spatial distance from the two-dimensional grid position to the deformation center position, a two-dimensional deformation probability distribution is generated using a Gaussian projective kernel function, and these distributions are stacked according to the sampling time order to form a spatial dimension consistent with the initial two-dimensional visual features in step S3. Figure 1 The initial two-dimensional spacetime deformation probability tensor.

[0014] Furthermore, in step S2, the edge computing unit forms a deformation shock jump value based on the temporal physical deformation scalar, forms a frequency domain residual significance based on the logarithmic amplitude spectrum of the video stream frame and the local mean filter array, and inputs the deformation shock jump value and the frequency domain residual significance into the support vector data description model to obtain the abnormal coupling deviation. Then, the feature mask is determined based on the comparison result between the abnormal coupling deviation and the preset safety threshold.

[0015] Furthermore, in step S2, when the abnormal coupling deviation is lower than the preset safety threshold, the edge computing unit confirms the initial two-dimensional spatiotemporal deformation probability tensor as the feature mask; when the abnormal coupling deviation reaches or exceeds the preset safety threshold, the edge computing unit outputs the fully transparent identity matrix as the feature mask, and the fully transparent identity matrix is ​​called in the form of an all-1 mask tensor in pixel-by-pixel feature modulation.

[0016] Furthermore, in step S3, the edge computing unit inputs the video stream frame into the visual feature extraction network to obtain an initial two-dimensional visual feature map, and the output layer of the visual feature extraction network retains the two-dimensional spatial position; the edge computing unit performs spatial size matching and channel direction copying on the feature mask according to the downsampling ratio of the video stream frame to the initial two-dimensional visual feature map, and then performs pixel-by-pixel feature modulation with the initial two-dimensional visual feature map to form a coupled feature map.

[0017] Furthermore, in step S4, the edge computing unit inputs the coupled feature map into the structural fracture feature extraction network, extracts the edge abrupt response by using directional gradient convolution kernels along the axial and circumferential directions of the target flexible pipe, and aggregates them into crack candidate regions according to the 8-neighborhood connection rule. Then, based on the linear continuity of the crack candidate regions, the consistency of local edge abrupt changes, and the consistency of cross-channel response, the crack identification and localization results are generated.

[0018] Furthermore, in step S4, the edge computing unit performs a fast Fourier transform on the temporal physical deformation scalar and eliminates the DC component to determine the energy main frequency, and compares the energy main frequency with the static water dead zone threshold; when the energy main frequency is lower than the static water dead zone threshold, the link execution state is dormant signaling. During the sliding window period of the dormant signaling, step S2 directly outputs the fully transparent identity matrix as a feature mask, and step S3 continues to perform pixel-by-pixel feature modulation.

[0019] The technical effects and advantages of the artificial intelligence-based crack identification method for flexible marine hydrogen transport pipelines of this invention are as follows: This invention binds spatial pose data, pipe centerline sampling points, and video stream frames for processing, so that the crack identification process no longer relies solely on the apparent texture in the video stream frames. After the inverse kinematics compensation operator deducts the apparent offset caused by the underwater inspection platform's own motion, the temporal physical deformation scalar can characterize the macroscopic bending curvature and transient tensile variables of the target flexible pipe, reducing the interference of underwater inspection platform attitude changes and position drift on pipe deformation judgment.

[0020] This invention maps the temporal physical deformation scalar to an initial two-dimensional spatiotemporal deformation probability tensor using a Gaussian projective kernel function, and generates a feature mask using deformation impact jump values, frequency domain residual significance, and a support vector data description model. This process allows both the pipe deformation state and the video image frequency domain state to participate in mask verification. When the abnormal coupling deviation reaches or exceeds a preset safety threshold, a fully transparent identity matrix is ​​used as the feature mask, maintaining the unit transparency relationship of the initial two-dimensional visual feature map and reducing the risk of the abnormal mask suppressing the actual crack response. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0022] Figure 2 This is a flowchart of the feature mask generation and coupling reliability assessment process of the present invention; Figure 3 This is a schematic diagram of pixel-by-pixel feature modulation and coupled feature map generation according to the present invention; Figure 4 This is a flowchart of the crack identification, localization, and link execution status control of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 - Figure 4 This invention provides an artificial intelligence-based method for identifying cracks in flexible marine hydrogen transport pipelines, comprising: This invention takes video stream frames, spatial pose data, and centerline sampling points of the target flexible pipeline as input. First, it binds the geometric reference of the pipeline in the video stream frames to the spatial pose data of the underwater inspection platform. Then, it subtracts the apparent offset caused by the underwater inspection platform's own motion using an inverse kinematics compensation operator, and organizes the remaining relative motion of the target flexible pipeline into a temporal physical deformation scalar. Subsequently, it maps the temporal physical deformation scalar to an initial two-dimensional spatiotemporal deformation probability tensor using a Gaussian projective kernel function, and simultaneously calculates the deformation impact jump value and frequency domain residual significance. Finally, it outputs the abnormal coupling deviation through a support vector data description model. In the visual processing stage, the edge computing unit extracts the initial two-dimensional visual feature map of the video stream frame, and modulates the feature mask and the initial two-dimensional visual feature map pixel by pixel according to sampling, spatial and channel position to form a coupled feature map. In the result generation stage, the edge computing unit extracts structural fracture features based on the coupled feature map and forms a structural fracture feature map, generates crack identification and positioning results, and outputs dormant signaling or complete link execution status based on the comparison between the energy main frequency of the temporal physical deformation scalar and the static water dead zone threshold.

[0025] S1: In response to the inspection command, acquire video stream frames and spatial pose data containing the target flexible pipe, extract the sampling points of the pipe centerline, and use the inverse kinematics compensation operator to compensate and generate a time-domain physical deformation scalar.

[0026] Specifically, step S1 includes the following: The following steps are all executed by the edge computing unit mounted on the underwater inspection platform. After responding to the inspection command, the edge computing unit acquires the video stream frame containing the target flexible pipe and its spatial pose data, and organizes them onto the same sampling time axis. Since the apparent motion of the target flexible pipe in the video stream frame includes both the underwater inspection platform's own swimming trajectory and the pipe deformation induced by ocean currents, step S1 uses an inverse kinematics compensation operator to deduct the apparent offset caused by the underwater inspection platform's own motion and generates a temporal physical deformation scalar that can be called by steps S2 and S4.

[0027] S101: Video stream frames, spatial pose data and tube centerline sampling points are synchronized.

[0028] Upon receiving the inspection command, the edge computing unit activates the machine vision module and the inertial navigation pose stream acquisition channel. The machine vision module continuously outputs video stream frames containing the target flexible pipe, while the inertial navigation pose stream synchronously outputs spatial pose data. The video stream frames are a sequence of image frames arranged in chronological order of acquisition; the spatial pose data are a continuous sampling sequence of 3D position, attitude quaternions, and Euler angle attitude information corresponding to the acquisition time of the video stream frames.

[0029] Edge computing units use the inspection command trigger time as a common time reference to determine the first... The video stream frames and spatial pose data at each sampling time are written to the same sampling node, forming a frame pose synchronization node. Each frame pose synchronization node contains at least the first sampling time frame. Each video stream frame, the corresponding pose quaternion, Euler angles, position vector, and sampling point of the tube centerline.

[0030] The sampling points for the pipe centerline are two-dimensional image coordinate points formed by continuously connecting the midpoints of the pipe edges along the axial direction of the target flexible pipe. The edge computing unit extracts the two sides of the target flexible pipe within the same video stream frame, calculates the midpoints of the two side edge points at the same axial position, and connects multiple midpoints sequentially along the axial direction of the target flexible pipe to form a sequence of sampling points for the pipe centerline. This sequence of sampling points is bound to the spatial pose data at the corresponding sampling time and used for subsequent inverse kinematics compensation and temporal physical deformation scalar calculations; it is not output as an independent crack identification result.

[0031] In one specific embodiment, when the underwater inspection platform inspects the outer wall of the target flexible pipe, the machine vision module sequentially captures images of the textured areas on the outer surface of the pipe, while the inertial navigation pose stream simultaneously records the attitude sway of the underwater inspection platform caused by thruster corrections and ocean current disturbances. The edge computing unit extracts the pipe centerline sampling point within each video stream frame and binds it to the spatial pose data at the same sampling time, enabling the apparent displacement of the pipe in the image to establish a correspondence with the underwater inspection platform's own motion state.

[0032] S102: Construction of inverse kinematics compensation operator.

[0033] After obtaining the frame pose synchronization node, the edge computing unit uses the spatial pose data as the starting point for processing, converts the pose quaternion at each sampling moment into a rotation matrix, and uses Euler angles to perform consistency verification on the rotation direction. The rotation matrix is ​​used to characterize the attitude change of the underwater inspection platform's body coordinate system relative to the global reference coordinate system, which is the reference coordinate system determined by the inertial navigation pose flow when the inspection command is triggered.

[0034] For the At each sampling time, the edge computing unit constructs a rigid body homogeneous transformation matrix based on the position vector and rotation matrix in the spatial pose data. This rigid body homogeneous transformation matrix describes the absolute motion state of the underwater inspection platform in the global reference coordinate system. Subsequently, the edge computing unit performs matrix inversion on the rigid body homogeneous transformation matrix to obtain the inverse homogeneous transformation matrix corresponding to that sampling time. The inverse homogeneous transformation matrices at consecutive sampling times are arranged in chronological order to form the inverse kinematics compensation operator.

[0035] The inverse homogeneous transformation matrix can be represented as: ; in, For the first The inverse homogeneous transformation matrix at each sampling time is used to subtract the coordinate changes caused by the underwater inspection platform's own motion; For the first Each sampling moment is a rotation matrix transformed from attitude quaternions and checked by Euler angles, used to characterize the attitude orientation of the underwater inspection platform; This is the transpose of the rotation matrix, used to perform the reverse rotation; For the first The position vector at each sampling moment is derived from spatial pose data and is used to characterize the absolute spatial displacement of the underwater inspection platform. This is a reverse translation term, used to offset the apparent offset caused by the displacement of the underwater inspection platform itself.

[0036] The inverse kinematics compensation operator does not generate the target flexible pipe deformation solely from spatial pose data. Instead, it applies the inverse homogeneous transformation matrix corresponding to the spatial pose data to the pipe centerline sampling points in the video stream frame at the same sampling time to subtract the apparent offset caused by the underwater inspection platform's own motion. Before the inverse homogeneous transformation matrix is ​​applied to the pipe centerline sampling points, the edge computing unit calls the camera intrinsic parameter matrix obtained by the machine vision module during the deployment phase and converts the pipe centerline sampling points from two-dimensional coordinates of the video stream frame to normalized camera coordinates. For the ... For any sampling point along the centerline of the pipe at any sampling time, the edge computing unit determines the depth scale of the sampling point based on the calibrated pipe diameter or inspection distance scale of the target flexible pipe. It then combines the normalized camera coordinates with the depth scale to form a three-dimensional spatial point, which is then written into a four-dimensional homogeneous coordinate form. The four-dimensional homogeneous coordinates serve as the input to the inverse homogeneous transformation matrix, used to subtract the apparent offset caused by the underwater inspection platform's own motion.

[0037] S103: Extraction of the relative motion state sequence of the target flexible pipe.

[0038] After the inverse kinematics compensation operator is formed, the edge computing unit will... The sampling points of the pipe centerline at each sampling time are converted into four-dimensional homogeneous coordinates, and the inverse homogeneous transformation matrix corresponding to that sampling time is called to deduct the apparent offset of the sampling points of the pipe centerline caused by the attitude and position changes of the underwater inspection platform itself.

[0039] After completing node-by-node compensation, the edge computing units compare the remaining displacement and remaining rotation angle of the sampling points along the pipe centerline at adjacent sampling times to form a relative motion state sequence of the target flexible pipe. The relative motion state sequence of the target flexible pipe is a continuous time series formed after processing by the inverse kinematics compensation operator, used to characterize the relative displacement, relative rotation angle, and their continuous changing trend of the target flexible pipe caused by ocean currents in a three-dimensional flow field.

[0040] In one specific embodiment, when the underwater inspection platform experiences a short-term yaw due to a lateral ocean current, the edge of the target flexible pipe in the video stream frame will show a synchronous shift. The edge computing unit uses the inverse homogeneous transformation matrix at the same sampling time to subtract the change in observation direction caused by the yaw; the pipe bending, axial expansion or contraction, or local sway that still exist after subtraction are written into the relative motion state sequence of the target flexible pipe.

[0041] S104: Generation of time-domain physical deformation scalars.

[0042] After obtaining the relative motion state sequence of the target flexible pipeline, the edge computing unit calculates the set of macroscopic bending curvature and transient stretching variables of the target flexible pipeline along the sampling time axis, and organizes the two into a one-dimensional continuous time series according to the sampling time to generate a time-domain physical deformation scalar.

[0043] Let the first The sampling points along the centerline of the pipe at each sampling time are represented in axial order according to the target flexible pipe. ,in, The sampling point number is the centerline sampling point number of the pipe body. Let be the coordinates of the sampling point in the two-dimensional grid of the video stream frame. The sampling time of the first sampling moment The macroscopic curvature of each centerline segment can be expressed as: ; in, For the first The sampling time of the first sampling moment The macroscopic curvature of each centerline segment is used to construct the time-domain physical deformation scalar. For the first The sampling time of the first sampling moment The orientation angle of each centerline segment is derived from the direction of the line connecting the sampling points of adjacent pipe centerlines; The change in the directional angle between adjacent centerline segments is used to characterize the degree of bending of the target flexible pipeline; For the first The sampling time of the first sampling moment The arc length interval of each centerline segment is derived from the distance between sampling points of adjacent pipe centerlines and is used to convert the change in direction angle into a curvature quantity.

[0044] No. The sampling time of the first sampling moment The transient stretching variable of each centerline segment can be expressed as: ; in, For the first The sampling time of the first sampling moment The transient stretching variables of each centerline segment are used to characterize the degree of local axial expansion and contraction of the target flexible pipe. For the current sampling time, the first... The length of each centerline segment is derived from the sampling points of the tube's centerline in the current video stream frame; The length of the same centerline segment at the previous sampling time is derived from the tube centerline sampling point in the previous video stream frame; This represents the change in segment length between adjacent sampling times, used to form the tensile response.

[0045] At each sampling moment, the edge computing unit calculates the macroscopic curvature and transient tensile variables of each centerline segment, and forms a deformation set for the current sampling moment according to the axial order of the target flexible pipe. Subsequently, the edge computing unit extracts the maximum absolute value of the macroscopic curvature and the average absolute value of the transient tensile variables from the deformation set at the current sampling moment, and normalizes them according to the corresponding maximum values ​​in the static water calibration samples. The normalized maximum absolute value of the macroscopic curvature is then added to the normalized average absolute value of the transient tensile variables to obtain the equivalent deformation scalar for that sampling moment. The edge computing unit arranges the equivalent deformation scalars for each sampling moment according to the sampling time order to form a time-domain physical deformation scalar. The time-domain physical deformation scalar serves as the physical output object of step S1, proceeds to step S2 for Gaussian projective kernel function mapping and deformation impact jump value calculation, and then proceeds to step S4 for energy frequency calculation.

[0046] S2: Receives the temporal physical deformation scalar and video stream frames, generates an initial two-dimensional spatiotemporal deformation probability tensor through the Gaussian projective kernel function, and performs a coupled reliability assessment based on the temporal physical deformation scalar and video stream frames to determine the feature mask.

[0047] Specifically, step S2 includes the following: Step S1 has already output the video stream frame and temporal physical deformation scalar bound to the same sampling time axis. Step S2 continues processing based on the above objects, mapping the temporal physical deformation scalar to an initial two-dimensional spatiotemporal deformation probability tensor covering the visual receptive field, and simultaneously calculating the deformation shock jump value and frequency domain residual significance. Then, it outputs the abnormal coupling deviation degree through the support vector data description SVDD model. The edge computing unit determines the feature mask that can be directly called in step S3 based on the comparison result of the abnormal coupling deviation degree and the preset safety threshold.

[0048] S201: Generate the initial two-dimensional spatiotemporal deformation probability tensor based on the Gaussian projective kernel function.

[0049] The edge computing unit reads the temporal physical deformation scalar output in step S1, and reads the video stream frame size and the spatial size of the subsequent initial two-dimensional visual feature map within the same sliding window period. If the two-dimensional deformation probability distribution is first generated in the video stream frame spatial size, the edge computing unit maps the two-dimensional deformation probability distribution to the initial two-dimensional visual feature map space according to the downsampling ratio from the video stream frame to the initial two-dimensional visual feature map; the final spatial size of the initial two-dimensional spatiotemporal deformation probability tensor is based on the spatial size of the initial two-dimensional visual feature map.

[0050] The edge computing unit first determines the first sampling point of the pipe centerline obtained in S1. The deformation center position at each sampling moment. The deformation center position is the location of the sampling point on the centerline of the tube where the macroscopic curvature reaches its maximum value at the current sampling moment, within the two-dimensional grid of the video stream frame. For any position in the two-dimensional grid... The edge computing unit calculates the Euclidean distance from the location to the deformation center and uses this Euclidean distance as the spatial decay input of the Gaussian projective kernel function.

[0051] The mapping relationship of the Gaussian projective kernel function can be expressed as: ; in, For the first Each sampling time is located at a position on the two-dimensional grid. The deformation probability value at that location is used to generate the initial two-dimensional spatiotemporal deformation probability tensor; For the first The normalized equivalent deformation scalar at each sampling time is derived from the time-domain physical deformation scalar generated in step S104 and normalized according to the maximum value of the deformation scalar in the static water calibration sample. It is used to characterize the amplitude of the deformation probability distribution at the current sampling time. The position of the two-dimensional grid is derived from the spatial coordinates corresponding to the video stream frame or the initial two-dimensional visual feature map. For the first The deformation center position at each sampling moment is derived from the sampling point of the tube centerline where the macroscopic curvature reaches the maximum value at the current sampling moment. The squared Euclidean distance from the two-dimensional grid position to the deformation center position is used to characterize the spatial attenuation of the deformation effect in the visual receptive field. The stress attenuation scale parameter for polymer materials is determined by the equivalent image distance corresponding to the stress response of the target flexible pipe material sample segment decreasing to a preset proportion of the peak response during bending loading calibration; in one embodiment, the preset proportion is... Edge computing units reduce the stress response to the peak response. The corresponding image equivalent distance is used as .

[0052] After all sampling moments within the sliding window period have undergone Gaussian projective kernel mapping, the edge computing units stack the two-dimensional deformation probability distributions of each sampling moment in chronological order to generate an initial two-dimensional spatiotemporal deformation probability tensor. The temporal dimension of the initial two-dimensional spatiotemporal deformation probability tensor is consistent with the sampling nodes of the temporal physical deformation scalar, and the spatial dimension is consistent with the spatial dimension of the initial two-dimensional visual feature map.

[0053] S202: Calculation of deformation impact jump value and frequency domain residual significance.

[0054] After the initial two-dimensional spatiotemporal deformation probability tensor is formed, the edge computing unit continues to use the temporal physical deformation scalar and video stream frames within the same sliding window period to calculate the reliability evaluation quantity. The edge computing unit reads the temporal physical deformation scalar along the sampling time axis retained in step S1, and reads the corresponding video stream frames according to the same sliding window period, so that the dynamic evaluation quantity and the optical evaluation quantity are kept in time consistent.

[0055] For the time-domain physical deformation scalar, the edge computing unit performs three consecutive time derivatives on the one-dimensional continuous time series within the sliding window period to obtain the third-order time derivative sequence. The root mean square of the third-order time derivative sequence is then calculated, and the deformation impact jerk value is output. The deformation impact jerk value is used to quantify the transient dynamic failure energy of a target flexible pipeline subjected to a sudden tearing load.

[0056] The deformation impact jump value can be expressed as: ; in, The deformation impact jump value is used as the input to the support vector data description SVDD model. The number of sampling nodes for the temporal physical deformation scalar within the sliding window period is derived from the sampling time axis retained in step S1; For the first The temporal physical deformation scalar at each sampling time; For the first The third-order time derivative value corresponding to each sampling time is used to characterize the degree of abrupt change in the rate of deformation.

[0057] When the time-domain physical deformation scalar is stored in the form of discrete sampling nodes, the edge computing unit approximates the third-order time derivative with the third-order difference between adjacent sampling nodes: ; in, The time interval between adjacent sampling nodes is derived from the sampling time axis retained in step S1; , , and These are time-domain physical deformation scalars for four consecutive sampling times, used to calculate approximate values ​​of the third-order time derivative in edge computing units.

[0058] For each video stream frame, the edge computing unit performs a Fast Discrete Fourier Transform (FFT) on the video stream frames within a sliding window period and extracts the logarithmic amplitude spectrum. Subsequently, the edge computing unit performs local mean filtering on the logarithmic amplitude spectrum, generating a local mean filter array. The local mean filter array is obtained by sliding the logarithmic amplitude spectrum through a fixed-size mean filter window. The size of the mean filter window is determined by rounding down the smaller of the video stream frame's width and height according to a preset ratio, and the window is supplemented at the boundary positions using a mirror expansion method. In one embodiment, the preset ratio is 1 / 32. When the mean filter window size obtained by rounding down according to the preset ratio is less than 3 spectral sampling points, the edge computing unit sets the mean filter window size to 3 spectral sampling points and performs local mean filtering using an odd-sized window.

[0059] The significance of frequency domain residuals can be expressed as: ; in, The frequency domain residual significance is used as the input to the support vector data description SVDD model. The number of video stream frames participating in the calculation within the current sliding window period is derived from the sliding window period formed in step S1; For the first The logarithmic amplitude spectrum obtained by performing a fast discrete Fourier transform on each video stream frame; For the first The local mean filter array is formed by applying local mean filtering to the logarithmic magnitude spectrum of each video stream frame; is the Frobenius norm of the difference matrix, used to characterize the frequency domain residual intensity of the video stream frame.

[0060] S203: Support Vector Data Description (SVDD) model output abnormal coupling deviation.

[0061] After both the deformation impact jump value and the frequency domain residual significance are generated, the edge computing unit organizes them into a current state input vector in a fixed order. The current state input vector uses the deformation impact jump value as the first component and the frequency domain residual significance as the second component, which is used to simultaneously characterize the transient dynamic destruction energy of the target flexible pipe and the frequency domain structural fidelity of the video stream frame.

[0062] In one embodiment, the Support Vector Data Description (SVDD) model uses state input vectors formed from historical sliding window periods as training samples. These state input vectors consist of deformation impact jump values ​​and frequency domain residual significance in a fixed order. The training dataset is taken from historical data collected from the target flexible pipeline under normal inspection conditions. Normal inspection conditions include sliding window periods without manually marked cracks, without pipe body obstruction, where the target flexible pipeline contour is continuous in the video stream frames, and where no frames are lost in the inertial navigation pose flow. When establishing the training dataset, the edge computing unit first divides the historical sliding window periods according to the sampling time axis of the video stream frames. Then, manual verification and labeling remove sliding window periods containing cracked areas, pipe body obstruction areas, interrupted target flexible pipeline contour areas, and missing spatial pose data. The remaining sliding window periods are written into the normal inspection condition sample set. The training dataset contains 12,000 historical sliding window cycle samples, of which 9,600 samples are used as the training set and 2,400 samples are used as the validation set. The deformation impact jump value and frequency domain residual significance of each sample are calculated in the same way as in step S202, and mean normalization is performed before inputting them into the model. The normalized mean and scale parameters are fixed and stored after being statistically analyzed from the training set. The Support Vector Data Description (SVDD) model is implemented as a kernelized anomaly detection computation graph. This computation graph includes an input layer, a normalization layer, a radial basis function (RBF) kernel mapping layer, and a hypersphere detection layer. The input layer has two input nodes, receiving the deformation shock jerk value and the frequency domain residual significance, respectively. The normalization layer has two normalization nodes, used to call the normalized mean and normalized scale parameters fixed in the training set. The RBF kernel mapping layer includes 128 support vector units, each using the RBF kernel function as its activation function. The initial value of the kernel width parameter is set to 0.35, and before training, it is updated to the fixed kernel width parameter based on the reciprocal of the median of the pairwise Euclidean distances between the historical state input vectors in the training set. The hypersphere detection layer includes one distance output node, used to output the anomaly coupling deviation of the current state point relative to the center of the smallest closed hypersphere. During training, the edge computing unit or offline training server uses the loss function of minimizing the sum of the squared hypersphere radius and the penalty term for out-of-bounds samples, with a penalty coefficient set to 0.05, a maximum number of iterations set to 2000, and a convergence tolerance set to 0.0001. After training, the support vector set, radial basis kernel width parameter, hypersphere center, boundary radius, normalized mean, and normalized scale parameter are fixed and saved; among them, the boundary radius is used as a preset safety threshold. During inspection, the edge computing unit does not update the above pre-trained parameters, but only inputs the current state input vector generated in the current sliding window cycle into the fixed support vector data description SVDD model to calculate the abnormal coupling deviation.

[0063] The radial basis function kernel function can be expressed as: ; in, The radial basis kernel function used to describe the SVDD model for support vector data; and For any two historical state input vectors, each vector is composed of deformation impact jump values ​​and frequency domain residual significance in a fixed order; The kernel width parameter is fixed, with an initial value of 0.35. It is updated before training based on the reciprocal of the median of the pairwise Euclidean distance between the input vectors of the historical states in the training set. This parameter is used to control the tightness of the sample envelope after high-dimensional mapping. The squared Euclidean distance is the distance between two historical state input vectors.

[0064] Abnormal coupling deviation can be expressed as: ; in, This represents the abnormal coupling deviation, used for comparison with a preset safety threshold. The input vector is the current state. The high-dimensional state point is obtained by mapping the current state input vector through the kernel function of the Support Vector Data Description (SVDD) model. The deformation impact jump value output by S202; The significance of the frequency domain residual output by S202; The center of the smallest closed hypersphere formed by the support vector data description SVDD model in the regenerating kernel Hilbert high-dimensional space; It is a Euclidean metric used to calculate the distance from the current state point to the center of the sphere.

[0065] Before being input into the Support Vector Data Description (SVDD) model, the current state input vector is normalized to the same scale as the normal operating data. The mean and scale parameters used for normalization are calculated separately from the deformation impact jump value sequence and the frequency domain residual significance sequence in the normal operating data and are fixed after model training. The preset safety threshold is fixedly stored in the edge computing unit after the SVDD model training is completed and is not updated with the input value of the current sliding window cycle during the inspection process.

[0066] S204: Feature mask confirmation and output.

[0067] After comparing the abnormal coupling deviation with the preset safety threshold, the edge computing unit enters the feature mask confirmation process. The feature mask is a spatial weight object output from step S2 and called by step S3.

[0068] When the abnormal coupling deviation is lower than a preset safety threshold, the edge computing unit recognizes the initial two-dimensional spatiotemporal deformation probability tensor as a feature mask. In this path, the feature mask retains the spatial probability distribution and sampling time order of the initial two-dimensional spatiotemporal deformation probability tensor, which is used for pixel-by-pixel feature modulation in step S3 with the initial two-dimensional visual feature map.

[0069] When the abnormal coupling deviation reaches or exceeds a preset safety threshold, the edge computing unit outputs a fully transparent identity matrix as a feature mask. This fully transparent identity matrix is ​​invoked as a full-one mask tensor in pixel-by-pixel feature modulation, and its spatial dimensions are the same as the initial two-dimensional visual features. Figure 1 For all spatial and channel locations, the mask weight is 1; when described in matrix form, this matrix is ​​used to represent the unit pass-through relationship, rather than assigning values ​​only at the two-dimensional diagonal positions.

[0070] Therefore, the feature mask output by step S2 is either an initial two-dimensional spatiotemporal deformation probability tensor that has been reliably verified, or a fully transparent identity matrix, and serves as the direct input for pixel-by-pixel feature modulation in step S3.

[0071] S3: Extract the initial two-dimensional visual feature map of the video stream frame, perform pixel-by-pixel feature modulation on the feature mask and the initial two-dimensional visual feature map in the spatial and channel dimensions, and output the coupled feature map.

[0072] Specifically, step S3 includes the following: Step S2 has already output the feature mask, and the video stream frames output in step S1 still maintain the original sampling timeline. Step S3 extracts the initial two-dimensional visual feature map based on the above objects, and applies the feature mask to the initial two-dimensional visual feature map according to the same sampling time, the same spatial position, and the same channel position to generate a coupled feature map for use in step S4.

[0073] S301: Initial two-dimensional visual feature map extraction.

[0074] The edge computing unit reads the video stream frames output in step S1 according to the sampling time sequence, and then... The video stream frames at each sampling time point are input into the visual feature extraction network. The initial two-dimensional visual feature map is a two-dimensional spatial feature response map formed after the video stream frames are processed by the visual feature extraction network. It is used to characterize the activation distribution of the surface texture, edge gradient, and local fracture texture of the target flexible pipe in the image space.

[0075] The visual feature extraction network consists of convolutional layers and non-linear activation layers in sequence. Its output layer retains the two-dimensional spatial location and does not perform global pooling operation that compresses the spatial location into a single classification vector. This allows the initial two-dimensional visual feature map to perform pixel-by-pixel feature modulation with the feature mask at the same spatial location.

[0076] In practical implementation, when the machine vision module captures the outer wall of the target flexible pipe, the axial texture, circumferential boundary, and local scratches of the pipe in the video stream frame are all entered into the visual feature extraction network. If the pipe produces texture ghosting due to ocean current swaying in the same frame, this texture ghosting will also form a local high response in the initial two-dimensional visual feature map.

[0077] S302: Spatiotemporal alignment of the feature mask with the initial 2D visual feature map.

[0078] After the initial two-dimensional visual feature map is formed, the edge computing unit calls the feature mask output in step S2 and establishes a spatial correspondence between it and the initial two-dimensional visual feature map at the same sampling time. There are two sources for the feature mask: when step S2 determines that the abnormal coupling deviation is below a preset safety threshold, the feature mask is the initial two-dimensional spatiotemporal deformation probability tensor; when step S2 determines that the abnormal coupling deviation reaches or exceeds the preset safety threshold, the feature mask is a fully transparent identity matrix.

[0079] The edge computing unit reads the data from the feature mask based on the sampling time axis. A two-dimensional mask slice corresponding to an initial two-dimensional visual feature map. If the spatial size of the feature mask is already consistent with the initial two-dimensional visual feature map... Figure 1 If the edge computing unit directly reads the mask weights at the corresponding positions, and if the feature mask still maintains the spatial size of the video stream frame, the edge computing unit performs position mapping according to the downsampling ratio of the video stream frame to the initial two-dimensional visual feature map before reading the mask weights.

[0080] When the initial 2D visual feature map contains multiple channels, the edge computing unit replicates the 2D mask slices with spatially matched dimensions along the channel direction, so that each channel obtains the same spatial mask weights. This channel replication only changes the multiplicative shape of the feature mask in the channel dimension, without changing the deformation probability value or unit transmission value of each spatial location in the feature mask.

[0081] S303: Pixel-wise feature modulation based on Hadamard product.

[0082] After completing spatiotemporal alignment, the edge computing unit is in the first... Taking an initial two-dimensional visual feature map and its corresponding feature mask as input, pixel-wise feature modulation is performed. Pixel-wise feature modulation involves using the feature mask output from step S2 to perform element-wise multiplication operations with the initial two-dimensional visual feature map in both channel and spatial dimensions, so that the visual activation value at each spatial location is preserved, suppressed, or transmitted according to the corresponding mask weights.

[0083] The operational relationship of pixel-wise feature modulation can be expressed as: ; in, For the first At each sampling time in spatial location and channels The response value of the coupled feature map on it is used as the output object of step S3; For the first At each sampling time in spatial location and channels The initial two-dimensional visual feature map response value is derived from S301; For the first At each sampling time in spatial location and channels The feature mask weights are derived from the feature mask after channel replication is completed by S302. This indicates element-wise multiplication at the same sampling time, the same spatial location, and the same channel location. When this operation is performed on the entire feature map, the element-wise multiplication is equivalent to the Hadamard product multiplication operation.

[0084] When the feature mask originates from the initial two-dimensional spatiotemporal deformation probability tensor, the edge computing unit modulates the activation values ​​of the corresponding spatial locations in the initial two-dimensional visual feature map according to the deformation probability values. For regions where high-frequency gradients are generated due to elastic deformation, the physical probability weights in the feature mask suppress the activation values ​​of interfering pixels in that region; for regions not covered by the deformation probability, the structural break activations in the initial two-dimensional visual feature map are retained.

[0085] When the feature mask originates from a fully transparent identity matrix, the edge computing unit expands it into a full-1 mask tensor of the same size as the initial 2D visual feature map and performs the same pixel-by-pixel multiplication operation, so that the initial 2D visual feature map remains transparent in all spatial and channel positions.

[0086] S304: Output of coupled feature map.

[0087] After pixel-by-pixel feature modulation is completed at each sampling moment within the sliding window period, the edge computing unit organizes the modulation results of each sampling moment according to the sampling time sequence to form a coupled feature map. The coupled feature map is a visual-physical coupled spatial feature map formed by pixel-by-pixel feature modulation of the feature mask and the initial two-dimensional visual feature map. It is used to characterize the feature response after the combined effect of the surface structure fracture activation and physical deformation prior of the target flexible pipe.

[0088] When the edge computing unit outputs the coupled feature map, it retains the sampling time axis and spatial position relationship consistent with the video stream frame, so that step S4 can read the coupled feature map at the same time node and extract the structural fracture features under the same spatial coordinates.

[0089] S4: Extract structural fracture features based on the coupled feature map and generate crack identification and location results. Output the link execution status based on the energy main frequency of the time-domain physical deformation scalar and the static water dead zone threshold.

[0090] Specifically, step S4 includes the following: Step S3 has already output a coupled feature map that inherits the sampling time axis and spatial position relationship of the video stream frames. The temporal physical deformation scalar formed in step S1 is still retained as a physical state input. Step S4 extracts structural fracture features from the coupled feature map and generates crack identification and localization results. At the same time, it outputs the link execution status based on the frequency domain characteristics of the temporal physical deformation scalar.

[0091] S401: Extraction of structural fracture features.

[0092] The edge computing unit reads the coupled feature map within the same sliding window period according to the sampling time axis, and then... The coupled feature map at each sampling time is input into the structural fracture feature extraction network. The structural fracture features are linear discontinuous edges, local gray-scale abrupt response, and cross-channel consistent fracture response extracted from the coupled feature map, which are used to characterize the spatial fracture morphology of suspected crack areas on the surface of the target flexible pipe.

[0093] The structural fracture feature extraction network comprises a set of oriented gradient convolution kernels and a connected component response aggregation layer. The oriented gradient convolution kernel set includes at least gradient convolution kernels along the axial direction of the target flexible pipe and along the circumferential direction of the target flexible pipe, used to extract edge abrupt responses along the pipe's extension direction and around the pipe's circumference, respectively. The connected component response aggregation layer first determines the crack response threshold based on the 90th quantile of the ascending order of the structural fracture response sequence of normal pipe samples. Then, spatial locations in the output of the oriented gradient convolution kernel set that are above the crack response threshold are designated as high-response locations and aggregated into candidate connected regions according to the 8-neighborhood connection rule. The length, width, principal direction, minimum bounding rectangle, center point coordinates, and average response across channels are recorded for each candidate connected region.

[0094] When the underwater inspection platform passes through the bend of the target flexible pipe, step S3 has already suppressed the texture ghosting that occurs synchronously with the elastic deformation or preserved it according to the full transmission path. The edge computing unit continues to search for linear continuous fracture responses on the coupled feature map; if there are thin cracks on the outer wall of the pipe, the directional gradient convolution kernel group forms adjacent high responses on both sides of the crack boundary, and the connected domain response aggregation layer aggregates this continuous linear region into a crack candidate region.

[0095] S402: Crack candidate region determination and crack identification and localization results generation.

[0096] After the structural fracture feature map is formed, the edge computing unit generates crack candidate regions using the structural fracture feature map as input. The crack candidate regions are spatial regions in the structural fracture feature map that satisfy the continuous linear response condition, and are used as the spatial source of crack identification and localization results.

[0097] The edge computing unit first sorts the response intensity of the structural fracture feature maps, and uses the 90th percentile of the ascending order of the structural fracture response sequences of normal pipe samples as the crack response threshold. Normal pipe samples are the structural fracture feature maps corresponding to historical video stream frames that have been manually verified and show no cracked areas. The crack response threshold is calculated and stored in the edge computing unit before deployment. The edge computing unit marks spatial locations in the structural fracture feature maps that are above the crack response threshold as fracture response points, and then performs 8-neighborhood connected component aggregation on adjacent fracture response points to form crack candidate regions.

[0098] For each crack candidate region, the edge computing unit calculates its minimum bounding rectangle, center point coordinates, extension direction along the pipe body, aspect ratio, and cross-channel response overlap ratio, and binds the above spatial positioning information to the corresponding sampling time. Linear continuity is determined by the aspect ratio of the minimum bounding rectangle of the crack candidate region. When the long side of the minimum bounding rectangle extends along the pipe body surface and the aspect ratio is higher than the preset aspect ratio threshold, the linear continuity requirement is met. The consistency of local edge abrupt changes is determined by the opposite signs and adjacent positions of the gradient responses on both sides of the crack candidate region. The consistency of cross-channel response is determined by the overlap ratio of the response positions of the same crack candidate region in multiple channels. When the overlap ratio of the response positions is higher than the preset overlap ratio threshold, the cross-channel response consistency requirement is met.

[0099] Both the preset aspect ratio threshold and the preset overlap ratio threshold are obtained statistically from the structural fracture feature maps corresponding to crack-free historical video stream frames and manually marked crack samples. Specifically, the preset aspect ratio threshold is the lower quartile of the ascending order of the aspect ratio sequence of the manually marked crack samples, and the preset overlap ratio threshold is the lower quartile of the ascending order of the channel overlap ratio sequence of the manually marked crack samples. Crack candidate regions that satisfy linear continuity, consistency of local edge abrupt changes, and consistency of cross-channel response are written into the crack identification and localization results; other crack candidate regions are only retained as intermediate responses in the current sliding window cycle.

[0100] The crack identification and localization result consists of the position of the crack region in the corresponding spatial coordinates of the video stream frame and the crack identification marker. When the spatial position of the crack region in the structural fracture feature map is inconsistent with the spatial size of the video stream frame, the edge computing unit maps the minimum bounding rectangle and center point coordinates of the crack region back to the video stream frame coordinates according to the upsampling ratio from the initial two-dimensional visual feature map to the video stream frame, and writes the mapped spatial coordinates and the crack identification marker together into the crack identification and localization result.

[0101] S403: Calculation of the energy dominant frequency of time-domain physical deformation scalars.

[0102] Simultaneously with the generation of crack identification and location results, the edge computing unit calls the time-domain physical deformation scalar output in step S1 and calculates the frequency domain characteristics along the same sliding window period. The dominant energy frequency is the frequency of the maximum non-DC energy component in the spectral energy distribution after the time-domain physical deformation scalar undergoes a fast Fourier transform, and is used to characterize the main flow-induced vibration state of the target flexible pipe within the current sliding window period.

[0103] The edge computing unit assembles the temporal physical deformation scalar within the sliding window period into a one-dimensional input sequence according to the sampling time order, and then inputs this one-dimensional input sequence into the Fast Fourier Transform (FFT) component to obtain the deformation spectrum sequence. Subsequently, the edge computing unit excludes the DC component with a frequency of 0, and selects the frequency component with the largest energy amplitude from the remaining frequency components as the dominant energy frequency.

[0104] The energy frequency can be represented as: ; in, This is the main energy frequency, used for determining the link's execution status. The time-domain physical deformation scalar output in step S1 is used to form a one-dimensional input sequence according to the sliding window period; The time-domain physical deformation scalar is transformed by the Fast Fourier Transform at the frequency. spectral components at the location; For frequency The energy amplitude at that point is used to determine the vibrational frequency component with the largest energy within the current sliding window cycle; This indicates the non-zero frequency range after excluding the DC component.

[0105] The static water area dead zone threshold is determined by static water area calibration samples. These samples are collected under conditions where the target flexible pipeline does not exhibit visible swaying and the underwater inspection platform maintains a stable cruising attitude. During collection, the time-domain physical deformation scalar is generated according to step S1, and the energy frequency is calculated in the same manner as in S403. The edge computing unit sorts the energy frequencies corresponding to multiple static water area calibration samples in ascending order and takes the 90th percentile as the static water area dead zone threshold. The static water area dead zone threshold is stored permanently in the edge computing unit after calibration and is not updated with the energy frequency of the current sliding window cycle during a single inspection.

[0106] S404: Link execution status determination and output.

[0107] After obtaining the energy master frequency and the static water dead zone threshold, the edge computing unit performs a link execution state determination. The link execution state is the result of control step S2 performing complete nonlinear feature mapping and mask reliability assessment in subsequent sliding window cycles, or entering the pure native visual direct recognition mode.

[0108] When the dominant energy frequency is lower than the static water dead zone threshold, the edge computing unit generates a dormant signal. This dormant signal acts on the execution control entry point of step S2 in the subsequent sliding window cycle, causing step S2 to skip the initial two-dimensional spatiotemporal deformation probability tensor generation, deformation impact jump value calculation, frequency domain residual significance calculation, and support vector data description (SVDD) model joint analysis, and directly output a fully transparent identity matrix as a feature mask. Step S3 continues to perform pixel-by-pixel feature modulation, ensuring the initial two-dimensional visual feature map remains transparent under the action of the all-1 mask tensor.

[0109] When the energy frequency reaches or exceeds the static water dead zone threshold, the edge computing unit outputs the complete link execution status. The complete link execution status acts on the execution control entry of step S2 in the subsequent sliding window cycle, causing step S2 to continue executing the Gaussian projective kernel function mapping of the time-domain physical deformation scalar, the calculation of the deformation impact jump value, the calculation of the frequency domain residual significance, the joint analysis of the support vector data description SVDD model, and the output of the feature mask.

[0110] Through the above processing, step S4 completes the conversion from the coupled feature map to the crack identification and localization result, and forms the link execution state based on the frequency domain characteristics of the time-domain physical deformation scalar. The crack identification and localization result provides the spatial location of the crack region on the surface of the target flexible pipe and the crack identification marker; the link execution state controls the execution path of step S2 in the subsequent sliding window cycle through sleep signaling or full link execution state, so that crack identification and computing power scheduling are closed on the same sampling time axis.

[0111] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements 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 method for identifying cracks in flexible marine hydrogen transport pipelines based on artificial intelligence, characterized in that, Includes the following steps: Step S1: In response to the inspection command, acquire video stream frames and spatial pose data containing the target flexible pipe, extract the sampling points of the pipe centerline, and use the inverse kinematics compensation operator to compensate and generate a time-domain physical deformation scalar. Step S2: Receive the temporal physical deformation scalar and video stream frames, generate an initial two-dimensional spatiotemporal deformation probability tensor through the Gaussian projective kernel function, and perform a coupling reliability assessment based on the temporal physical deformation scalar and video stream frames to determine the feature mask; Step S3: Extract the initial two-dimensional visual feature map of the video stream frame, perform pixel-by-pixel feature modulation on the feature mask and the initial two-dimensional visual feature map in the spatial and channel dimensions, and output the coupled feature map; Step S4: Extract structural fracture features based on the coupled feature map and generate crack identification and location results, and output the link execution status based on the energy main frequency of the time-domain physical deformation scalar and the static water dead zone threshold.

2. The method for identifying cracks in a flexible marine hydrogen transport pipeline based on artificial intelligence according to claim 1, characterized in that, In step S1, the edge computing unit uses the inspection command trigger time as a common time reference, writes the video stream frame, spatial pose data and pipe centerline sampling point at the same sampling time into the frame pose synchronization node, and retains the sampling time axis; the pipe centerline sampling point is formed by connecting the midpoints of the two sides of the target flexible pipe at the same axial position in axial order, and inputs the inverse kinematics compensation operator.

3. The method for identifying cracks in a flexible marine hydrogen transport pipeline based on artificial intelligence according to claim 2, characterized in that, In step S1, the edge computing unit converts the attitude quaternions in the spatial pose data into rotation matrices, performs orientation verification using Euler angles, and then constructs a rigid body homogeneous transformation matrix by combining the position vector. The rigid body homogeneous transformation matrix is ​​inverted to form inverse kinematic compensation operators arranged in the order of sampling time. The inverse kinematic compensation operators are applied to the sampling points of the pipe centerline to deduct the apparent offset caused by the movement of the underwater inspection platform.

4. The method for identifying cracks in a flexible marine hydrogen transport pipeline based on artificial intelligence according to claim 3, characterized in that, In step S1, the edge computing unit calculates the change in direction angle and arc length interval of adjacent centerline segments based on the sampling points of the tube centerline after deducting the apparent offset by the inverse kinematics compensation operator to obtain the macroscopic bending curvature, and calculates the change in length of the same centerline segment at adjacent sampling times to obtain the transient stretching variable, and then generates the time-domain physical deformation scalar according to the sampling time sequence.

5. The method for identifying cracks in a flexible marine hydrogen transport pipeline based on artificial intelligence according to claim 4, characterized in that, In step S2, the edge computing unit takes the sampling point of the tube centerline where the macroscopic curvature reaches the maximum value at the current sampling time as the deformation center position. Based on the spatial distance from the two-dimensional grid position to the deformation center position, it generates a two-dimensional deformation probability distribution through the Gaussian projective kernel function and stacks them in the order of sampling time to form an initial two-dimensional spatiotemporal deformation probability tensor with the same spatial size as the initial two-dimensional visual feature map in step S3.

6. The method for identifying cracks in a flexible marine hydrogen transport pipeline based on artificial intelligence according to claim 5, characterized in that, In step S2, the edge computing unit forms a deformation shock jump value based on the temporal physical deformation scalar, forms a frequency domain residual significance based on the logarithmic amplitude spectrum of the video stream frame and the local mean filter array, and inputs the deformation shock jump value and the frequency domain residual significance into the support vector data description model to obtain the abnormal coupling deviation. Then, the feature mask is determined based on the comparison result between the abnormal coupling deviation and the preset safety threshold.

7. The method for identifying cracks in a flexible marine hydrogen transport pipeline based on artificial intelligence according to claim 6, characterized in that, In step S2, when the abnormal coupling deviation is lower than the preset safety threshold, the edge computing unit confirms the initial two-dimensional spatiotemporal deformation probability tensor as the feature mask. When the abnormal coupling deviation reaches or exceeds the preset safety threshold, the edge computing unit outputs a fully transparent identity matrix as a feature mask. The fully transparent identity matrix is ​​called in the form of an all-1 mask tensor in pixel-by-pixel feature modulation.

8. The method for identifying cracks in a flexible marine hydrogen transport pipeline based on artificial intelligence according to claim 7, characterized in that, In step S3, the edge computing unit inputs the video stream frame into the visual feature extraction network to obtain an initial two-dimensional visual feature map, and the output layer of the visual feature extraction network retains the two-dimensional spatial position. The edge computing unit performs spatial size matching and channel direction copying on the feature mask according to the downsampling ratio from the video stream frame to the initial two-dimensional visual feature map, and then performs pixel-by-pixel feature modulation with the initial two-dimensional visual feature map to form a coupled feature map.

9. The method for identifying cracks in a flexible marine hydrogen transport pipeline based on artificial intelligence according to claim 8, characterized in that, In step S4, the edge computing unit inputs the coupled feature map into the structural fracture feature extraction network, extracts the edge abrupt response by directional gradient convolution kernel groups along the axial and circumferential directions of the target flexible pipe, and aggregates them into crack candidate regions according to the 8-neighborhood connection rule. Then, based on the linear continuity of the crack candidate regions, the consistency of local edge abrupt changes, and the consistency of cross-channel response, the crack identification and localization results are generated.

10. The method for identifying cracks in a flexible marine hydrogen transport pipeline based on artificial intelligence according to claim 9, characterized in that, In step S4, the edge computing unit performs a fast Fourier transform on the temporal physical deformation scalar and eliminates the DC component to determine the energy main frequency. The energy main frequency is then compared with the static water dead zone threshold. When the energy main frequency is lower than the static water dead zone threshold, the link execution state is sleep signaling. During the sliding window period of the sleep signaling, step S2 directly outputs the fully transparent identity matrix as a feature mask, and step S3 continues to perform pixel-by-pixel feature modulation.

Images