Pipe defect discrimination method based on polarity flip characteristics of pulse eddy current differential signals

Abstract

The application relates to the field of nondestructive testing signal processing and provides a pipeline defect discrimination method based on polarity flip characteristics of a pulsed eddy current differential signal, which comprises the following steps: determining polarity flip events based on the pulsed eddy current differential signal, extracting node features reflecting local physical properties, determining edges of an event graph based on time causality between events and neighborhood signal form correlation degree; generating an event representation based on the event graph, determining physical state parameters; driving a physical state differential equation based on the physical state parameters, predicting the theoretical physical state of each event, and fusing the event representation to generate a fusion representation; and outputting a pipeline defect discrimination result based on the fusion representation. The polarity flip events are constructed into a graph structure with physical causal connection, and the event evolution reasoning process is constrained through a physical state differential equation, so that the robustness and interpretability of defect discrimination are improved.

Description

Technical Field

[0001] This application relates to the field of nondestructive testing signal processing, and in particular to a pipeline defect discrimination method based on the polarity reversal characteristics of pulsed eddy current differential signals. Background Technology

[0002] like Figure 4 As shown, pulsed eddy current detection is a common technique for detecting defects in conductive structures such as metal pipes and container walls. Conventional axial differential probes can acquire differential time-domain response signals, where the polarity reversal process contains information about the eddy current circulation and attenuation caused by pipe wall defects. Currently, common defect discrimination methods often treat the differential signal as a time-series waveform, image, or sequence feature, directly inputting it into a classifier or regression model. However, these methods struggle to fully express the physical causal relationships between polarity reversal events and lack explicit constraints on the electromagnetic physical processes such as the diffusion and attenuation of pulsed eddy currents within the pipe wall. When the detection conditions involve lift-off variations, local permeability fluctuations, or strong electromagnetic interference, the discrimination accuracy and robustness of the aforementioned conventional methods are easily and significantly affected, and the interpretability of the identification results is insufficient. Summary of the Invention

[0003] This application provides a pipeline defect discrimination method based on the polarity reversal characteristics of pulsed eddy current differential signals, which solves the technical problems in the prior art of insufficient expression of the physical causal relationship between polarity reversal events in differential signals and lack of constraints on the physical diffusion process of pulsed eddy currents.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a pipeline defect identification method based on the polarity reversal characteristics of pulsed eddy current differential signals is provided, including: Based on pulsed eddy current differential signals, multiple polarity reversal events are identified. Node features reflecting local physical properties are extracted for these multiple polarity reversal events. Based on the temporal causality and neighborhood signal morphology correlation between the polarity reversal events, the edges of the event graph are determined, thus obtaining the event graph. Based on the event graph, an event representation of each polarity reversal event is generated, and physical state parameters are determined; Based on the physical state parameters, a preset physical state differential equation is driven to predict the theoretical physical state corresponding to each polarity reversal event, and a fused representation is generated based on the theoretical physical state and the event representation. Based on the fusion characterization, the pipeline defect discrimination result is output.

[0005] Based on the above technical solution, in the pipeline defect discrimination method based on the polarity reversal characteristics of pulsed eddy current differential signals provided in this application, discrete polarity reversal events in pulsed eddy current differential signals are constructed into an event graph with physical attribute nodes and causal correlation edges. The physical parameters caused by the defect drive the physical state differential equation to predict the theoretical state at each event moment. Then, the physical prediction state is fused with the data-driven event representation, so that the defect discrimination process is explicitly guided by the laws of electromagnetic physics, thereby maintaining high discrimination accuracy and anti-interference ability under interference conditions.

[0006] In conjunction with the first aspect above, in one possible implementation, the node features include: The flip sharpness feature reflects the speed of the polarity flip event; The local attenuation coefficient characteristics reflect the eddy current attenuation characteristics of the signal segment where the polarity reversal event occurs; Symmetry characteristics reflect the asymmetry of the waveforms before and after the polarity reversal event; Multi-scale energy characteristics reflect the complexity of signal patterns near the polarity reversal event.

[0007] By selecting multiple node features that reflect the physical essence of the flipping process, each node in the event graph is characterized from the perspectives of instantaneous change rate, energy dissipation, spatial asymmetry, and multi-scale complexity. This allows the event graph to more completely carry the disturbance information of defects on the physical process of pulsed eddy currents, providing a highly discriminative node representation basis for subsequent graph reasoning and physical consistency fusion.

[0008] In conjunction with the first aspect above, in one possible implementation, the edges of the event graph are determined in the following way: Based on the temporal causality between the polarity reversal events, the existence and / or temporal causal strength of edges are determined; based on the correlation of neighborhood signal morphology, the edge weights and / or connection strengths are determined; wherein, the edge weights comprehensively characterize the temporal causality and the correlation of neighborhood signal morphology.

[0009] By combining the two dimensions of temporal causality and morphological similarity to determine the existence and strength of edges in an event graph, the edges can not only reflect the apparent similarity between signal segments, but also reflect the cause-and-effect relationship between events caused by eddy current circulation and attenuation. This results in a graph topology structure that is more in line with the physical logic of electromagnetic induction, so that the subsequent graph message transmission process follows the true causal order.

[0010] In conjunction with the first aspect above, in one possible implementation, the physical state parameters include at least the remaining wall thickness-related parameters and the material permeability-related parameters. The physical state differential equation uses the state vector related to the magnetic flux density or magnetic vector potential as the state variable. The coefficient terms of the physical state differential equation are determined based on the physical state parameters, and the theoretical physical state of subsequent polarity reversal events is predicted using the state corresponding to the first polarity reversal event as the initial value.

[0011] The residual wall thickness parameter and material property parameter, which can reflect the nature of the pipe wall defect, are explicitly introduced into the coefficient terms of the physical state differential equation, so that the evolution of the equation is completely dominated by the physical parameters of the defect. Taking the state of the first actual detected flip event as the starting point of the evolution, the operating point that should be at the subsequent event time is predicted. The theoretical state sequence can characterize the ideal magnetic flux flip behavior under given defect conditions, and provide an independent physical reference for subsequent data characterization.

[0012] In conjunction with the first aspect above, in one possible implementation, the flip sharpness feature is determined based on a combination of the instantaneous frequency at the polarity flip event and the derivative of the analytic signal envelope; The local attenuation coefficient characteristics are obtained by performing fractional-order exponential damping fitting on the signal segments segmented by the polarity reversal event, resulting in the attenuation order and relaxation time. The symmetry feature is determined based on a distance metric between the energy distribution of the signal segments before and after the flip; The multi-scale energy features are based on wavelet decomposition with a variable quality factor, which yields multiple sub-band energy entropies and compresses them into low-dimensional features.

[0013] By integrating various signal processing techniques such as instantaneous frequency analysis, fractional-order damping fitting, distribution distance measurement, and quality factor variable wavelet transform, node features can be extracted. This allows for high-precision reconstruction of the local behavior of eddy current physical response from multiple dimensions, including time-frequency joint characteristics, multi-scale attenuation mechanism, statistical distribution asymmetry, and frequency band complexity. This enables node features of different defect types to have better separability in the feature space.

[0014] In conjunction with the first aspect above, in one possible implementation, the temporal causality between the polarity reversal events is determined based on the inter-event propagation entropy or a temporal causal window; the morphological correlation of the neighborhood signal is determined based on the curvature complexity of the derivative dynamic time warping path of the neighborhood signal of the polarity reversal event. The edge weights satisfy: ;in, Let be the edge weight between node i and node j. The temporal causality strength from node i to node j Let be the curvature complexity of the derivative of the signal in the neighborhood of the polarity reversal event, representing the dynamic time-warped path. This is the scaling parameter.

[0015] The edge weights are determined by a combination of directed causal measures such as transfer entropy and morphological measures based on DTW path curvature complexity. This allows the edge weights to not only distinguish between defect-related flipping modes and random noise flipping modes, but also to amplify the information interaction between nodes with strong causal correlation and high morphological consistency in graph message transmission, suppress crosstalk of irrelevant paths, and improve the signal-to-noise ratio of event graph representation.

[0016] In conjunction with the first aspect above, in one possible implementation, the state vector is: ,in, The magnetic vector potential is in cylindrical coordinates; The physical state differential equation is:

[0017] in, For parameters related to the remaining wall thickness, These are parameters related to the magnetic permeability of the material. The term represents the square of the system's natural frequency, determined by the remaining wall thickness and permeability. The damping coefficient term is determined by the remaining wall thickness and magnetic permeability. The excitation coupling coefficient is determined by the remaining wall thickness. This is an excitation signal.

[0018] The transient physical process of pulsed eddy currents in the pipe wall is abstracted into a second-order vector potential diffusion equation with the magnetic vector potential and its rate of change as the core state. The natural frequency, damping, and coupling coefficient of the equation are all analytically determined by the physical parameters of the defect. This equation can generate a theoretical flip state sequence that conforms to the laws of electromagnetics in a rigorous derivation manner, providing a highly accurate reference benchmark for physics-data fusion and significantly reducing the dependence of pure data-driven models on the number of training samples and the coverage of operating conditions.

[0019] In conjunction with the first aspect above, in one possible implementation, the step of generating a fusion representation based on the theoretical physical state and the event representation includes: using the theoretical physical state as a query and the event representation as a key and value; generating attention weights based on the matching relationship between the theoretical physical state and the event representation; and aggregating the event representation based on the attention weights to obtain calibration features, which serve as the fusion representation.

[0020] By designing the physical prediction state as a query in the attention mechanism, the data-driven event representation is recalibrated under the guidance of the physical state. The electromagnetic dynamic constraints implicit in the physical state can explicitly guide the model to focus on event features consistent with real physical behavior and extract the information most relevant to the current physical evolution state from the event features, thereby generating a fusion representation that is closer to the essence of the defect.

[0021] In conjunction with the first aspect above, in one possible implementation, the attention weights are obtained through optimal transport solving, including: Construct a cost matrix between the theoretical physical state and the event representation; solve the cost matrix based on entropy-regularized optimal transport to obtain the transport matrix, which is then used as attention weights.

[0022] The matching between physical states and event features is modeled as an optimal transport problem. The Sinkhorn algorithm with entropy regularization is used to obtain a transport matrix with row and column normalization and sparsity as attention weights. Compared with standard dot product attention, this mechanism can more globally and structurally align the feature distribution in the physical space and data space, suppress the interference of noise points, and thus improve the discriminative power of the fused representation while maintaining physical consistency.

[0023] In a second aspect, this application provides an electronic device, comprising: a processor; and a memory storing computer instructions that, when executed by the processor, implement the methods described in the first aspect and any possible implementation thereof.

[0024] This application provides a pipeline defect discrimination method based on the polarity reversal characteristics of pulsed eddy current differential signals. By representing the polarity reversal events in the pulsed eddy current differential signals as an event graph with physical attribute nodes and causal morphological correlation edges, the eddy current circulation and attenuation process caused by defects is mapped to the node features and edge topology in the graph structure. This not only preserves the physical information of local reversal events, but also characterizes the evolutionary correlation between events through temporal causality and morphological similarity. Thus, the model can obtain a more essential physical representation than directly processing time series or image signals, which is beneficial to maintaining the stability of defect discrimination under conditions such as lift-off changes and magnetic permeability fluctuations.

[0025] Meanwhile, by driving the physical state differential equation through defect physical parameters such as remaining wall thickness and magnetic permeability, the theoretical physical state corresponding to each flipping event is predicted using the magnetic vector potential dynamics model. Using this theoretical state as a query, the event representation is calibrated through cross attention or optimal transmission attention, thus realizing the organic combination of physical dynamic constraints and data-driven reasoning. This mechanism not only makes the final fused representation closer to the essence of the defect, but also makes the defect discrimination results have a traceable electromagnetic physical basis, significantly improving the interpretability of the method.

[0026] Furthermore, by integrating multi-dimensional physical properties such as flip sharpness, local attenuation characteristics, symmetry, and multi-scale energy entropy into node features, and by employing a joint measurement of temporal causal strength based on indicators such as transfer entropy and morphological correlation based on DTW curvature complexity in edge construction, and by introducing structured alignment methods such as optimal transmission in fusion calibration, the discriminant model's ability to decouple interference features and the fidelity of its representation of defect features are synergistically enhanced from three levels: feature extraction, topology construction, and characterization fusion. This results in an overall improvement in the defect identification accuracy and generalization ability of pulsed eddy current detection. Attached Figure Description

[0027] Figure 1 A schematic diagram of the overall process of the pipeline defect identification method provided in the embodiments of this application; Figure 2 A schematic diagram of theoretical physical state prediction based on physical state differential equations provided in this application embodiment; Figure 3 A schematic diagram of the optimal transmission fusion mechanism for theoretical physical states and event representations provided in the embodiments of this application; Figure 4 This is a schematic diagram of the pulsed eddy current detection principle in the embodiments of this application. Detailed Implementation

[0028] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0029] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0030] Example 1: like Figure 1As shown, this embodiment provides a pipeline defect identification method based on the polarity reversal characteristics of pulsed eddy current differential signals. This method can be executed by a computing device, such as an industrial computer, embedded processing board, edge computing node, or cloud server. The method includes steps S100 to S400.

[0031] Step S100: Based on the pulsed eddy current differential signal, determine multiple polarity reversal events, extract node features reflecting local physical properties for multiple polarity reversal events, and determine the edges of the event graph based on the temporal causality and the correlation degree of the neighborhood signal morphology between polarity reversal events to obtain the event graph.

[0032] The detection signal is a response signal acquired from the object under test using pulsed eddy current detection technology. The object under test can be, for example, a pipe, container wall, plate, or other conductive structure. The detection signal can be a differential time-domain signal from an axial differential probe, or a single-channel or multi-channel time-domain signal acquired by other types of probes.

[0033] It should be noted that the detection signal can be acquired by reading it in real time from the acquisition hardware, or by reading pre-acquired signal data from the storage medium or data bus.

[0034] The term "polarity reversal event" generally refers to a specific moment or local process in the detection signal that reflects a change in the direction of eddy current induction or a local reversal of the magnetic state, and is not mechanically limited to a zero crossing in a strict mathematical sense.

[0035] For example, when the detected signal is directly represented as a differential signal, the polarity reversal event can correspond to the zero-crossing moment of the signal; when the detected signal has a bias or baseline drift, the polarity reversal event can also correspond to the moment when the signal crosses a preset baseline or an adaptive baseline; the polarity reversal event can also be located by detecting the extreme point of the derivative of the signal, the instantaneous phase jump point, etc. As long as the local time position of the polarity change can be identified, it can be used as the polarity reversal event referred to in this invention.

[0036] The aforementioned node features are information extracted to describe the local physical environment surrounding each polarity reversal event. These features reflect the transient physical properties of eddy currents from dimensions such as reversal rate, energy attenuation, waveform symmetry, and signal mode complexity, enabling each event node to carry rich electromagnetic response information. The specific types and extraction methods of node features will be further elaborated in subsequent embodiments.

[0037] It should be further explained that this embodiment considers two dimensions simultaneously when determining the edges of the event graph: First, the temporal causality between polarity reversal events reflects the cause-and-effect relationship between events caused by the diffusion and attenuation of pulsed eddies in the pipe wall. For example, one event may be caused by eddies circulating around or secondary induction caused by another event. Second, the correlation of neighboring signal morphology focuses on the similarity or correlation of signal waveforms near two polarity reversal events in terms of morphology, because the reversal modes caused by the same defect often have consistency in morphology.

[0038] By using a joint metric of these two dimensions to construct the edges of the event graph, the graph structure can simultaneously carry the physical essence of node attributes and the causal logic of edge connections. Compared to simple sequential connections or spatial adjacency, this graph structure can better capture the disturbance information of defects on the physical process of pulsed eddy currents, and enable subsequent graph message transmission processes to follow the true physical causal order, thus maintaining stable graph representation capabilities under disturbance conditions such as lift-off changes and permeability fluctuations.

[0039] Step S200: Based on the event graph, generate event representations for each polarity reversal event and determine the physical state parameters.

[0040] It should be noted that after obtaining the event graph consisting of polarity-flipping event nodes and physical causal edges, the event graph can be processed through a graph message passing network to generate an event representation for each polarity-flipping event.

[0041] Among them, event representation is a high-level abstraction of node features and their neighborhood context information in the graph, which integrates information from other events that have causal and morphological relationships with it; the specific structure of the graph message passing network can be flexibly selected, such as graph attention network, graph convolutional network, or message passing network with edge feature gating, etc.

[0042] In some implementations, physical state parameters may include parameters related to the remaining wall thickness and parameters related to the material's magnetic permeability, as well as other physical quantities that affect eddy current diffusion and attenuation; that is, physical state parameters refer to quantities that can describe the physical properties of defects that cause differences in the current detection signal.

[0043] For example, physical state parameters can be determined by mapping graph-level features obtained from event graph pooling through a parametric regression network, or they can be directly generated based on event graph representations.

[0044] Separating event-level representation from physical state parameters that describe the nature of defects enables the system to both perceive signal patterns from data and explain the causes of defects from a physical perspective, providing clear parameter entry points for subsequent physical constraints.

[0045] Step S300: Based on the physical state parameters, drive the preset physical state differential equation to predict the theoretical physical state corresponding to each polarity reversal event, and generate a fused representation based on the theoretical physical state and event representation.

[0046] The physical state differential equation is a pre-established dynamic model based on electromagnetic laws. The coefficients of this equation are determined by the physical state parameters; therefore, once the physical state parameters are determined, the equation can describe the ideal flux reversal behavior of the pulsed eddy current under the corresponding defect conditions. Using the state corresponding to the first polarity reversal event in the detection process as the initial value, the equation is numerically solved for each subsequent polarity reversal event, yielding a series of theoretical physical states. These theoretical states constitute a theoretical evolution sequence under given physical parameters. The specific form of the physical state differential equation will be elaborated in subsequent embodiments.

[0047] It should be noted that theoretical physical states represent ideal behavior conforming to physical laws, while event representations are data-driven features extracted from actual detected signals. Both describe different time segments of the same physical process from different perspectives. This step fuses them, using states predicted by physical laws to calibrate and guide the data-driven representation, ensuring that the final fused representation retains detailed information from the real signal while maintaining physical consistency. Specific fusion methods can employ attention mechanisms or optimal transmission, among other cross-spatial information alignment techniques, which will be described in more detail in subsequent embodiments.

[0048] The fused representation not only has stronger anti-interference ability, but also provides more interpretable feature basis for defect identification, because the significant feature dimensions often correspond to the actual physical parameter changes of the defect.

[0049] Step S400: Based on the fusion characterization, output the pipeline defect discrimination result.

[0050] Among them, the defect identification results can be specifically represented as defect classification labels, defect severity levels, estimated remaining wall thickness, defect location information, or corrosion degree assessment values, etc., and the specific form can be determined according to the actual application scenario.

[0051] For example, in a pipeline corrosion detection task, the output can be a classification result of "severe corrosion / mild corrosion / normal" or a regression result at the millimeter level of the remaining wall thickness. The fusion characterization obtained through the aforementioned steps carries deep defect information calibrated with physical constraints, enabling subsequent classifiers or regressors to output accurate discrimination results with a relatively simple structure. This helps reduce the computational cost and overfitting risk when the model is deployed in an industrial environment.

[0052] Through steps S100 to S400, this embodiment transforms the processing of pulsed eddy current detection signals into a pipeline of event graph construction, graph reasoning, physical differential equation prediction, and cross-modal fusion. This allows the defect discrimination results to benefit from data-driven pattern recognition capabilities while also being constrained and guided by physical laws, thus realizing a pipeline defect discrimination method with high robustness and interpretability.

[0053] Example 2: This embodiment, based on Embodiment 1, provides a detailed explanation of the specific types and extraction methods of node features. In this embodiment, node features include at least two of the following four categories: flip sharpness features, local attenuation coefficient features, symmetry features, and multi-scale energy features.

[0054] In this embodiment, all four types of features are extracted to obtain a comprehensive characterization of the local physical properties of the polarity reversal event.

[0055] The polarity reversal sharpness feature reflects the speed of the polarity reversal event. Specifically, the Hilbert transform is applied to the detected signal to obtain the analytic signal, and the instantaneous frequency and envelope derivative are extracted at the reversal event point. The instantaneous frequency reflects the instantaneous speed of the reversal, and the envelope derivative reflects the drastic change in signal amplitude at the reversal moment. The reversal sharpness feature is determined based on the combination of this instantaneous frequency and the envelope derivative of the analytic signal. For example, a harmonic average of the two can be used to highlight cases where both are significant simultaneously, avoiding excessive influence of any outlier on the feature.

[0056] The local attenuation coefficient characteristic reflects the eddy current attenuation characteristics of the signal segment where the polarity reversal event occurs. This embodiment uses fractional-order exponential damping fitting on the signal segment segmented by the polarity reversal event. For example, a damping model incorporating the Mittag-Leffler function is used for nonlinear fitting to obtain the attenuation order and relaxation time. The attenuation order reflects the non-exponential nature of the attenuation process, and the relaxation time reflects the characteristic time scale of eddy current attenuation. Together, they constitute a two-dimensional local attenuation coefficient characteristic, which can more precisely characterize the multi-scale eddy current loss caused by anomalous diffusion than a single exponential attenuation constant.

[0057] Symmetry features reflect the asymmetry of the waveform before and after a polarity reversal event. Magnetic field torsion caused by defects often desymmetrs the signal waveform before and after the reversal event. This embodiment extracts signal segments of a certain length (e.g., one cycle length) before and after the reversal event, treats the normalized energy distributions of the two segments as two probability distributions, and calculates the distance metric between them as the symmetry feature. This distance metric can be, for example, the Wasserstein distance, which captures the overall difference in waveform shape better than simple area ratios or amplitude ratios, and has better robustness to noise and local distortions.

[0058] Multi-scale energy features reflect the complexity of signal patterns near polarity reversal events. Eddy current anomalies caused by defects may exhibit different pattern characteristics at different time scales. This embodiment employs wavelet transform with a variable quality factor (e.g., adjustable Q-factor wavelet transform) to perform multi-scale decomposition of the signal, obtaining multiple sub-bands under different quality factors. Energy entropy is calculated for each sub-band to form a multi-scale energy entropy matrix. Then, dimensionality reduction (e.g., principal component analysis) is used to compress the high-dimensional entropy matrix into low-dimensional features, preserving the nonlinear structure while reducing feature redundancy.

[0059] The four types of node features described above characterize the local physical properties of polarity reversal events from the perspectives of reversal dynamics, decay physics, waveform symmetry, and multi-scale mode complexity. Compared with traditional simple features that only rely on slope or amplitude, these features can provide richer and more physically discriminative node information for event graphs, thereby improving the accuracy of subsequent graph reasoning and defect detection.

[0060] Example 3: This embodiment, based on embodiment 1, provides a detailed explanation of how to determine the edges of the event graph.

[0061] In this embodiment, the determination of event graph edges is based on two dimensions: temporal causality and the correlation of neighborhood signal morphology. The edge weight comprehensively represents the contribution of both dimensions. First, based on the temporal causality between polarity reversal events, the existence and temporal causal strength of edges are determined. The diffusion and attenuation of pulsed eddies in the pipe wall have a clear temporal directionality; an earlier event may physically lead to a later event. This causal relationship can be measured using transfer entropy in information theory.

[0062] Specifically, for any two polarity reversal events, the propagation entropy from the neighborhood signal of event i to the neighborhood signal of event j is calculated. If the propagation entropy exceeds a preset threshold, it can be determined that there is a temporal causal edge from i to j, and the value of the propagation entropy is directly used as the temporal causal strength.

[0063] As an alternative, a time-causal window approach can be used, which involves establishing candidate edges based on event pairs that are temporally adjacent and whose time difference is within a preset window, and then filtering them based on the strength of time causality.

[0064] Secondly, edge weights are determined based on the morphological correlation of neighboring signals. This embodiment measures morphological correlation by calculating the curvature complexity of the derivative dynamic time warping path of the neighboring signals of two polarity reversal events. Derivative dynamic time warping aligns the trend of signal changes rather than absolute values, thus better capturing the essential morphological similarity of the reversal patterns. The lower the curvature complexity of the path, the smoother the alignment path of the two signal segments and the more similar their morphologies.

[0065] In some implementations, edge weights satisfy the following relationship: ,in The temporal causal strength (e.g., propagation entropy) from node i to node j. Let be the curvature complexity of the derivative of the signal in the neighborhood of the polarity reversal event, representing the dynamic time-warped path. The scaling parameter is used. This weighting form ensures that the higher the temporal causal strength and the lower the morphological complexity (i.e., the more similar the morphology), the larger the edge weight, indicating a stronger physical connection between the two events.

[0066] The joint metric approach enables the event graph to simultaneously reflect the temporal causal order of eddy current physical processes and the morphological consistency of defect-induced flipping patterns. Compared to graphing methods based solely on temporal adjacency or spatial distance, the event graph in this embodiment can more accurately capture the perturbation characteristics of defects on pulsed eddy current physical processes, providing a more reliable graph topology for subsequent graph message passing.

[0067] Example 4: like Figure 2 As shown, this embodiment, based on embodiment 1, provides a detailed explanation of the specific forms of the physical state parameters and the physical state differential equations.

[0068] The physical state parameters include at least the parameters related to the remaining wall thickness (denoted as...). ) and material permeability-related parameters (denoted as The remaining wall thickness directly affects the diffusion path length and attenuation rate of eddy currents within the pipe wall, while the permeability affects the coupling strength and diffusion coefficient of electromagnetic induction. These two parameters are the most important physical quantities affecting the pulsed eddy current response. The specific values ​​of the physical state parameters can be obtained by mapping the graph-level features after event graph pooling through a parametric regression network.

[0069] The physical state differential equation uses the state vector, which is related to the magnetic vector potential, as the state variable. In the tube wall eddy current problem, the magnetic vector potential can naturally describe the electromagnetic field distribution in the axisymmetric structure.

[0070] In some implementations, the state vector is taken as That is, the magnetic vector potential in cylindrical coordinates Its time derivative constitutes a second-order system. The physical state differential equation is a second-order state-space equation of the following form: .

[0071] in, The square term of the system's natural frequency, determined by the remaining wall thickness and permeability, reflects the inherent oscillation characteristics of the pipe wall in response to eddy currents. The damping coefficient term, determined by the remaining wall thickness and magnetic permeability, reflects the attenuation characteristics of the eddy current. The excitation coupling coefficient is determined by the remaining wall thickness, reflecting the coupling strength between the external excitation pulse and the eddy current in the tube wall; This is the excitation signal. These coefficient terms are all derived from the physical state parameters. The decision can be made in advance using analytical formulas or numerical simulations in electromagnetism.

[0072] The equation takes the state corresponding to the first polarity reversal event as its initial value and solves it using numerical integration methods (such as the Runge-Kutta method or the symplectic Euler method) at the time of each subsequent polarity reversal event, outputting a series of theoretical physical states. These theoretical states represent the state at a given defect physical parameters. The ideal evolution sequence of pulsed eddy current flux reversal behavior is given below. Since the equations are derived entirely from physical principles and do not contain learnable parameters, the predictions are physically consistent and interpretable.

[0073] It should be noted that the specific form of the physical state differential equation is not limited to the second-order equation based on the magnetic vector potential mentioned above.

[0074] In other implementations, a state vector related to magnetic flux density can be selected, or a state-space equation can be derived based on an equivalent RL circuit model, as long as the physical state parameters are used as coefficients, the state of the first flip event is used as the initial value, and the theoretical physical state of subsequent events can be predicted.

[0075] Example 5: like Figure 3 As shown, this embodiment, based on embodiment 1, provides a detailed explanation of the mechanism for generating fused representations based on theoretical physical states and event representations.

[0076] The generation of fusion representations employs a cross-modal attention mechanism.

[0077] In some implementations, the theoretical physical state is used as the query, and the event representation is used as the key and value.

[0078] This design is based on the following considerations: theoretical physical states represent the ideal behavior that should be exhibited under given physical parameters, and are therefore suitable as active retrievers to retrieve matching information from data-driven event representations; while event representations contain rich details from actual signals, making them suitable as a knowledge base to be retrieved. By calculating the matching relationship between the two, attention weights are generated, and then the event representations are weighted and aggregated based on the attention weights to obtain calibration features, which serve as the fused representation.

[0079] This process is equivalent to using the expectations of physical laws to calibrate and filter data-driven observations, so that the fused representation can both retain the detailed information in the real signal and remove noise or interference components that are inconsistent with physical laws.

[0080] In other implementations, attention weights are obtained through optimal transport. First, a cost matrix is ​​constructed between the theoretical physical state and the event representation, where each element represents the distance (e.g., the square of the Euclidean distance) between the i-th theoretical state and the j-th event representation. Then, this cost matrix is ​​solved based on entropy-regularized optimal transport: under the condition of row and column normalization constraints (i.e., transport quality conservation), the transport matrix that minimizes the total transport cost while maximizing entropy is found. This optimization problem can be solved iteratively using the Sinkhorn algorithm to obtain the transport matrix, which is then used as the attention weight matrix.

[0081] Compared to the standard softmax attention mechanism, the attention weights in the optimal transport solution have two advantages: first, natural sparsity, as optimal transport tends to concentrate quality on a few matching pairs; second, row and column normalization ensures global quality conservation, allowing each theoretical state to receive adequate information calibration. The entropy regularization term enhances the numerical stability and robustness of the solution.

[0082] The resulting calibration features, after fusion, constitute the fused representation, which simultaneously carries the consistency constraints of physical laws and data-driven details. This fused representation can then be pooled (e.g., by taking the mean of all calibration features or through statistical pooling) to obtain a graph-level representation for subsequent pipeline defect identification.

[0083] Example 6: This embodiment provides an electronic device for implementing the pipeline defect identification method in any of the foregoing embodiments. The electronic device includes a processor and a memory. The processor can be a central processing unit, graphics processing unit, neural network processor, embedded processor, or field-programmable gate array (FPGA), or a combination of such processors. The memory stores computer instructions, which, when executed by the processor, implement the pipeline defect identification method as described in Embodiment 1. The memory can include random access memory, read-only memory, flash memory, solid-state drive, or other non-volatile storage media.

[0084] In practical deployments, this electronic device can be an industrial computer, an embedded processing board, an edge computing node, or a cloud server. When deployed as an industrial computer or embedded processing board, the device can directly connect to the pulsed eddy current detection probe via I / O interfaces or communication modules to acquire detection signals in real time and perform defect identification. When deployed as an edge computing node or cloud server, the device can receive and process pre-collected detection signal data via a data bus or network. The various method steps can be executed by the same processor or distributed to different processors. For example, graph message passing and graph inference can be accelerated by a GPU or NPU, while the solution of the physical state differential equations can be performed by a CPU. This invention does not impose any limitations on this.

[0085] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0086] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A method for identifying pipeline defects based on the polarity reversal characteristics of pulsed eddy current differential signals, characterized in that, include: Based on pulsed eddy current differential signals, multiple polarity reversal events are identified. Node features reflecting local physical properties are extracted for these multiple polarity reversal events. Based on the temporal causality and neighborhood signal morphology correlation between the polarity reversal events, the edges of the event graph are determined, thus obtaining the event graph. Based on the event graph, an event representation of each polarity reversal event is generated, and physical state parameters are determined; Based on the physical state parameters, a preset physical state differential equation is driven to predict the theoretical physical state corresponding to each polarity reversal event, and a fused representation is generated based on the theoretical physical state and the event representation. Based on the fusion characterization, the pipeline defect discrimination result is output.

2. The method according to claim 1, characterized in that, The node features include at least two of the following categories: The flip sharpness feature reflects the speed of the polarity flip event; The local attenuation coefficient characteristics reflect the eddy current attenuation characteristics of the signal segment where the polarity reversal event occurs; Symmetry characteristics reflect the asymmetry of the waveforms before and after the polarity reversal event; Multi-scale energy characteristics reflect the complexity of signal patterns near the polarity reversal event.

3. The method according to claim 1, characterized in that, The edges of the event graph are determined as follows: Based on the temporal causality between the polarity reversal events, the existence and / or temporal causal strength of edges are determined; based on the correlation of neighborhood signal morphology, the edge weights and / or connection strengths are determined; wherein, the edge weights comprehensively characterize the temporal causality and the correlation of neighborhood signal morphology.

4. The method according to claim 1, characterized in that, The physical state parameters include at least the remaining wall thickness-related parameters and the material permeability-related parameters. The physical state differential equation uses the state vector related to the magnetic flux density or magnetic vector potential as the state variable. The coefficient terms of the physical state differential equation are determined based on the physical state parameters, and the theoretical physical state of subsequent polarity reversal events is predicted with the state corresponding to the first polarity reversal event as the initial value.

5. The method according to claim 2, characterized in that, The flip sharpness feature is determined based on a combination of the instantaneous frequency at the polarity flip event and the envelope derivative of the analytical signal; The local attenuation coefficient characteristics are obtained by performing fractional-order exponential damping fitting on the signal segments segmented by the polarity reversal event, thus obtaining the attenuation order and relaxation time. The symmetry feature is determined based on a distance metric between the energy distribution of the signal segments before and after the flip; The multi-scale energy features are based on wavelet decomposition with a variable quality factor, which yields multiple sub-band energy entropies and compresses them into low-dimensional features.

6. The method according to claim 3, characterized in that, The temporal causality between polarity reversal events is determined based on inter-event propagation entropy or a temporal causal window. The morphological correlation degree of the neighborhood signal is determined based on the curvature complexity of the derivative dynamic time warp path of the neighborhood signal of the polarity reversal event. The edge weights satisfy: ;in, Let be the edge weight between node i and node j. The temporal causality strength from node i to node j Let be the curvature complexity of the derivative of the signal in the neighborhood of the polarity reversal event, representing the dynamic time-warped path. This is the scaling parameter.

7. The method according to claim 4, characterized in that, The state vector is ,in, The magnetic vector potential is in cylindrical coordinates; The physical state differential equation is: In the formula, For parameters related to the remaining wall thickness, These are parameters related to the magnetic permeability of the material. The term represents the square of the system's natural frequency, determined by the remaining wall thickness and permeability. The damping coefficient term is determined by the remaining wall thickness and magnetic permeability. The excitation coupling coefficient is determined by the remaining wall thickness. This is an excitation signal.

8. The method according to claim 4, characterized in that, The generation of the fused representation based on the theoretical physical state and the event representation includes: Use the theoretical physical state as the query and the event representation as the key and value; Attention weights are generated based on the matching relationship between the theoretical physical state and the event representation; The event representations are aggregated based on the attention weights to obtain calibration features, which serve as the fused representations.

9. The method according to claim 8, characterized in that, The attention weights are obtained through optimal transfer, including: Construct the cost matrix between the theoretical physical state and the event representation; The cost matrix is ​​solved based on entropy-regularized optimal transport to obtain the transport matrix, which is then labeled as attention weights.

10. An electronic device, comprising: processor; A memory storing computer instructions that, when executed by the processor, implement the method according to claim 1.

Images