A Defect Detection Method for Railway Contact Wire Based on Eddy Current

Abstract

This invention belongs to the technical field of defect detection, specifically relating to a method for detecting defects in railway contact wires based on eddy currents. It addresses the technical problem that existing methods struggle to adapt to the dynamic changes in the geometric characteristics of the contact wire in different sections, leading to unstable signal decomposition results. The detection method includes: S1, generating a composite excitation signal; S2, mapping the response impedance signal and background magnetic field distribution data to an instantaneous surface coordinate system to obtain multimodal sensing data with a unified spatiotemporal reference; S3, simultaneously extracting the extreme values ​​of the local surface curvature as a third type of defect feature; and S4, constructing a probabilistic graphical model with the first, second, and third types of defect features as inputs and the defect type and severity level as outputs. This invention not only achieves the identification of defect types and levels but also quantifies the uncertainty of diagnostic results, improving the accuracy, robustness, and reliability of the detection.

Description

Technical Field

[0001] This invention belongs to the technical field of defect detection, specifically relating to a method for detecting defects in railway contact wires based on eddy currents. Background Technology

[0002] As a core component of the "pantograph-catenary" system in high-speed electrified railways, the health of the railway contact wire directly affects the power supply stability and operational safety of trains. With the continuous increase in train speed and density, the contact wire, under the combined effects of long-term mechanical friction, arc erosion, and environmental corrosion, will develop various defects such as wear, cracks, abrasions, and burns. If these defects are not detected and addressed in a timely manner, they can lead to pantograph-catenary failures, affecting train operation efficiency, or even catastrophic accidents such as wire breakage. Therefore, efficient and accurate non-destructive testing of railway contact wires is one of the core technologies for ensuring safe railway operation.

[0003] Currently, non-destructive testing technology for contact wires mainly relies on single-frequency eddy current testing, which typically identifies defects by detecting changes in the probe's response impedance. However, eddy current testing methods have the following limitations: 1) The defect information contained in the single-frequency excitation signal is limited in dimension, making it difficult to effectively distinguish and quantify defect types with complex morphologies and diverse causes; 2) On high-speed trains, the detection system is subjected to strong mechanical vibrations and complex electromagnetic interference, leading to a sharp drop in the signal-to-noise ratio, making weak defect signals easily submerged by noise; 3) Existing methods often analyze eddy current signals in isolation, ignoring the three-dimensional geometry of the contact wire itself (such as curvature changes and irregular wear) and the influence of background magnetic field fluctuations, making it prone to false alarms and missed alarms due to lift-off effects or external interference.

[0004] To improve detection accuracy and anti-interference capabilities, several improved techniques have been proposed. For example, frequency sweeping or multi-frequency eddy current techniques are used, combined with time-frequency analysis methods such as wavelet transform and empirical mode decomposition, and fused with sensor data from visual or laser 3D scanning. While these methods improve detection performance to some extent, several challenges remain: 1) Existing signal processing methods (such as empirical mode decomposition) typically rely on engineering experience for parameter selection, lacking adaptability and struggling to adapt to dynamic changes in the geometric features of the contact line in different sections. This leads to unstable signal decomposition results, and the loss or distortion of key defect feature information during processing. 2) The extracted features are mostly limited to shallow physical quantities such as time-domain or frequency-domain statistics (e.g., amplitude, phase, frequency). Their ability to represent deep singularities caused by early weak defects or complex defects (e.g., sudden changes in signal dynamics) is limited, resulting in low feature discrimination. 3) In the defect decision-making stage, fixed threshold discrimination methods or traditional pattern recognition methods are often used, making it difficult to integrate multi-source heterogeneous feature information and to quantitatively assess the uncertainty of detection results. The generalization ability and diagnostic reliability of the model need improvement.

[0005] Therefore, how to develop a system that can adaptively process signals, deeply mine the defect singularity features in multimodal data, and establish an intelligent diagnostic model capable of performing uncertainty reasoning is a technical challenge that urgently needs to be solved in the current field. Summary of the Invention

[0006] This invention provides a railway contact wire defect detection method based on eddy currents to solve the technical problem that existing methods are difficult to adapt to the dynamic changes in the geometric characteristics of the contact wire in different sections, resulting in unstable signal decomposition effects.

[0007] A method for detecting defects in railway contact wires based on eddy currents includes the following steps:

[0008] S1, use a pseudo-random binary sequence to perform phase modulation on a preset swept-frequency chirped baseband signal to generate a composite excitation signal, and use the composite excitation signal to drive the eddy current probe.

[0009] S2, synchronously acquires the response impedance signal of the eddy current probe, the three-dimensional geometric data of the contact line, and the background magnetic field distribution data in the circumferential direction of the eddy current probe; constructs an instantaneous surface coordinate system of the contact line based on the three-dimensional geometric data, and maps the response impedance signal and background magnetic field distribution data to the instantaneous surface coordinate system to obtain multimodal sensing data with unified spatiotemporal reference.

[0010] S3, calculate the local surface curvature of the three-dimensional geometric data, adaptively adjust the parameters of variational mode decomposition based on the local surface curvature, perform variational mode decomposition on the response impedance signal, and obtain a set of intrinsic mode functions; reconstruct the phase space of each obtained intrinsic mode function to construct an attractor, calculate the continuous homology of the attractor, and extract the Betti number sequence representing the singularity of the signal as the first type of defect feature; calculate the spatial gradient tensor of the background magnetic field distribution data and extract the maximum eigenvalue as the second type of defect feature, and at the same time, extract the extreme value of the local surface curvature as the third type of defect feature;

[0011] S4. Construct a probabilistic graphical model with the first, second, and third types of defect features as input and the defect type and severity level as output. Perform probabilistic inference through the probabilistic graphical model, calculate the posterior probability of defects at each detection location, and identify the detection locations with posterior probability of defects exceeding a preset threshold as defect points.

[0012] Furthermore, in S1, when generating the composite excitation signal, a linear feedback shift register of a specific order is used to generate an m-sequence, and the m-sequence is used as the pseudo-random binary sequence.

[0013] Obtain the start frequency, end frequency, and sweep period of the swept chirped baseband signal;

[0014] The composite excitation signal is obtained by using binary phase shift keying to modulate the swept-frequency chirped baseband signal with the m-sequence.

[0015] Furthermore, the linear feedback shift register is a 10th-order linear feedback shift register.

[0016] Furthermore, the start frequency of the swept-frequency chirped baseband signal is 5kHz, the end frequency is 50kHz, and the sweep period is 1ms.

[0017] Furthermore, in S2, when mapping the response impedance signal and background magnetic field distribution data to the instantaneous surface coordinate system, at each sampling time, the center line tangential vector of the laser scanning profile of the contact line is extracted as the U-axis, the normal vector of the wear surface of the contact line profile is extracted as the V-axis, and the W-axis is obtained by the cross product of the U-axis and the V-axis. The right-handed instantaneous surface coordinate system is constructed by the U-axis, V-axis and W-axis together.

[0018] At the corresponding sampling time, the response impedance signal of the eddy current probe is mapped to the origin of the right-hand instantaneous surface coordinate system, and the background magnetic field distribution data is mapped to the right-hand instantaneous surface coordinate system according to the sampling position.

[0019] Furthermore, in S3, when performing variational mode decomposition on the response impedance signal, when the absolute value of the local surface curvature is less than 0.01... At that time, the number of modes K in the variational mode decomposition is set to 4, and the penalty factor is... Set to 2000;

[0020] When the absolute value of the local surface curvature is between 0.01 Up to 0.05 When the interval is between, the number of modes K in the variational mode decomposition is set to 6, and the penalty factor is... Set to 1500;

[0021] When the absolute value of the local surface curvature is greater than 0.05 At that time, the number of modes K in the variational mode decomposition is set to 8, and the penalty factor is... Set to 1000.

[0022] Furthermore, in S3, when extracting the Betti number sequence representing the singularity of the signal, the average mutual information method is used to calculate the time delay parameters of each intrinsic mode function. Set the embedding dimension m to 5;

[0023] Based on time delay parameters Given an embedding dimension m, perform delayed embedding on the eigenmode functions to construct an attractor point cloud;

[0024] The 0th and 1st order continuous cohomology of the attractor point cloud were calculated based on the Vietoris-Rips complex, and the corresponding Betti numbers were extracted. and Sequence is used as a first-class defect feature.

[0025] Furthermore, in S3, when calculating the spatial gradient tensor of the background magnetic field distribution data and extracting the maximum eigenvalue, a Hall sensor arranged in a 3×3 array centered on the eddy current probe is used to simultaneously acquire triaxial background magnetic field vector data at nine spatial locations. ;

[0026] Using the central difference method, the spatial partial derivatives of the magnetic field strength in three orthogonal directions are calculated based on the triaxial background magnetic field vector data of nine spatial locations, forming the gradient tensor matrix. :

[0027] ;

[0028] For the gradient tensor matrix Perform eigenvalue decomposition and extract the eigenvalue with the largest absolute value as the second type of defect feature.

[0029] Furthermore, in S4, when constructing the probabilistic graphical model, a linear chain conditional random field is used as the probabilistic graphical model;

[0030] The continuous measurement points along the contact line detection path are regarded as a sequence. The first, second and third types of defect features corresponding to each measurement point in the sequence together constitute the observation value of the probabilistic graphical model.

[0031] The defect type is set as three discrete states: "surface crack", "internal crack" and "severe wear", and the severity level is set as three discrete states: "slight", "moderate" and "severe". These nine combined states are used as the output labels of the probabilistic graphical model.

[0032] The probabilistic graphical model is trained using a dataset of known defective samples.

[0033] Furthermore, the number of known defective samples is 500.

[0034] The beneficial effects are as follows: This invention uses a phase-modulated swept-frequency chirped composite signal as the excitation source, enhancing the information carrying capacity and anti-interference performance of the original signal, laying the foundation for subsequent refined analysis. By simultaneously acquiring eddy current response, three-dimensional geometry, and background magnetic field data, and mapping and fusing them under a unified instantaneous surface coordinate system, complete multimodal sensing data is constructed, which can suppress the lift-off effect caused by contact line geometric fluctuations and interference from the external electromagnetic environment. This invention further utilizes topological data analysis methods to extract the Betti number sequence representing the intrinsic dynamic singularity of defects from the response signal. Combined with multi-source information such as background magnetic field gradient and surface curvature extrema, a more discriminative multi-dimensional feature set is constructed, resulting in stronger identification capabilities for early, weak defects. Finally, by constructing a probabilistic graphical model to fuse and probabilistically infer multi-source heterogeneous features, not only is the identification of defect type and level achieved, but the uncertainty of diagnostic results can also be quantitatively evaluated, improving the accuracy, robustness, and reliability of detection. Attached Figure Description

[0035] Figure 1 This is a flowchart of a railway contact wire defect detection method based on eddy currents.

[0036] Figure 2 This is a schematic diagram illustrating the generation of a composite excitation signal.

[0037] Figure 3 This is a schematic diagram illustrating the adaptive adjustment of variational mode decomposition parameters based on curvature.

[0038] Figure 4 This is a schematic diagram of multimodal feature extraction. Detailed Implementation

[0039] An embodiment of the railway contact wire defect detection method based on eddy current provided by this invention:

[0040] like Figure 1 As shown, a method for detecting defects in railway contact wires based on eddy currents includes the following steps:

[0041] S1 uses a pseudo-random binary sequence to perform phase modulation on a preset swept-frequency chirped baseband signal to generate a composite excitation signal, which is then used to drive the eddy current probe.

[0042] Specifically, first, the parameters of the linearly swept-frequency chirped baseband signal are determined, such as a start frequency of 1kHz, an end frequency of 100kHz, and a duration of 10ms. Then, an m-sequence is generated as a pseudo-random binary sequence using a 10-stage linear feedback shift register. Next, the 0s and 1s in this pseudo-random binary sequence are mapped to -1 and +1, and the swept-frequency chirped baseband signal is phase-modulated using binary phase-shift keying: the original signal phase is maintained during the time period with a sequence value of +1, and the signal phase is reversed by 180° during the time period with a sequence value of -1, thus generating a composite excitation signal. Finally, the composite excitation signal is converted into an analog voltage signal by a digital-to-analog converter, amplified by a power amplifier, and input to the excitation coil of the eddy current probe to complete the probe drive, as shown below. Figure 2 As shown.

[0043] In an optional embodiment, in S1, when generating the composite excitation signal, a linear feedback shift register of a specific order is used to generate an m-sequence, and the m-sequence is used as the pseudo-random binary sequence.

[0044] Obtain the start frequency, end frequency, and sweep period of the swept chirped baseband signal;

[0045] The composite excitation signal is obtained by using binary phase shift keying to modulate the swept-frequency chirped baseband signal with the m-sequence.

[0046] For example, the linear feedback shift register uses a 10th-order linear feedback shift register, with a start frequency of 5kHz, an end frequency of 50kHz, and a sweep period of 1ms for the swept chirped baseband signal.

[0047] The first step is to generate the basic signals: First, two basic signals are generated. Using a 10th-order linear feedback shift register, an m-sequence with a period length of 1023 is generated based on the register feedback logic. This m-sequence possesses good pseudo-random characteristics. Simultaneously, a linear frequency modulated (LFM) chirped signal is generated. The frequency of this LFM chirped signal smoothly and linearly increases from 5kHz to 50kHz within a 1ms time period, serving as the baseband signal carrying the information.

[0048] The second step is composite phase modulation: using the m-sequence as the modulation signal, binary phase shift keying is performed on the linear frequency modulated (LFM) chirped baseband signal. Specifically, when the code element of the m-sequence is 0, the phase of the LFM chirped baseband signal remains unchanged; when the code element of the m-sequence is 1, the phase of the LFM chirped baseband signal is flipped by 180°. Through this modulation method, the resulting composite excitation signal not only possesses the broadband characteristics of a swept-frequency signal but also incorporates the pseudo-random phase coding of the m-sequence, enhancing the signal's anti-interference capability and detection resolution.

[0049] S2 synchronously acquires the response impedance signal of the eddy current probe, the three-dimensional geometric data of the contact line, and the background magnetic field distribution data in the circumferential direction of the eddy current probe; based on the three-dimensional geometric data, it constructs an instantaneous curved surface coordinate system of the contact line, and maps the response impedance signal and the background magnetic field distribution data to the instantaneous curved surface coordinate system to obtain multimodal sensing data with unified spatiotemporal reference.

[0050] Specifically, a high-precision clock source, such as a GPS second pulse signal, is used to simultaneously trigger the eddy current signal acquisition card, the line laser profile scanner, and the Hall sensor array arranged around the probe to begin synchronous data acquisition. The eddy current signal acquisition card records the voltage and current of the probe's pickup coil for calculating the response impedance; the line laser profile scanner acquires three-dimensional point cloud data of the contact line detection area; and the Hall sensor array measures the magnetic field components in three orthogonal directions around the probe. For a detection point at a certain moment, the three-dimensional point cloud data of that detection point is used to fit the contact line profile wear, and an instantaneous surface coordinate system is constructed: the tangential vector of the contact line is extracted as the U-axis, the normal vector of the profile wear is extracted as the V-axis, and the W-axis is obtained by the cross product of the U-axis and V-axis, together constructing a right-handed instantaneous surface coordinate system that moves with the detection point. The eddy current response impedance value at this moment is mapped to the origin of the instantaneous surface coordinate system. At the same time, according to the physical installation position of each Hall sensor, the measured magnetic field vector data is transformed into this new right-handed instantaneous surface coordinate system through coordinate transformation, completing the spatiotemporal reference unification of multimodal data.

[0051] In an optional embodiment, in S2, when mapping the response impedance signal and background magnetic field distribution data to the instantaneous surface coordinate system, at each sampling time, the center line tangential vector of the contact line laser scanning profile is extracted as the U-axis, the normal vector of the wear surface of the contact line profile is extracted as the V-axis, and the W-axis is obtained by the cross product of the U-axis and the V-axis. The right-hand instantaneous surface coordinate system is constructed by the U-axis, V-axis and W-axis.

[0052] At the corresponding sampling time, the response impedance signal of the eddy current probe is mapped to the origin of the right-hand instantaneous surface coordinate system, and the background magnetic field distribution data is mapped to the right-hand instantaneous surface coordinate system according to the sampling position.

[0053] Step 1: Dynamic Definition of the Local Coordinate System. First, establish a local coordinate system that dynamically changes with the detection position. For the 3D contour data of the contact line acquired by the detection system at each sampling moment, calculate the tangent direction of the center line of the contour at that position, and define the unit vector of this direction as the U-axis of the coordinate system. Simultaneously, calculate the normal direction of the wear surface of the contour at that detection position (i.e., the direction perpendicular to the surface), and define the unit vector of this direction as the V-axis.

[0054] Step 2: Coordinate System Refinement and Multimodal Data Mapping. The coordinate system is refined and the data mapped. By calculating the cross product of the U-axis and V-axis vectors, a new vector perpendicular to both the U and V axes is obtained. This new vector is defined as the W-axis. The U, V, and W axes together form a complete right-handed instantaneous surface coordinate system. The response impedance values ​​measured by the eddy current probe at the sampling time are precisely mapped to the origin of this new coordinate system. Simultaneously, based on the actual spatial position of the background magnetic field sensor array, the collected magnetic field distribution data is transformed into this local coordinate system through coordinate transformation, achieving alignment of multiple measurement data on a unified reference closely related to the surface morphology of the contact line.

[0055] S3, calculate the local surface curvature of the three-dimensional geometric data, adaptively adjust the parameters of variational mode decomposition based on the local surface curvature, perform variational mode decomposition on the response impedance signal, and obtain a set of intrinsic mode functions; reconstruct the phase space of each obtained intrinsic mode function to construct an attractor, calculate the continuous homology of the attractor, and extract the Betti number sequence representing the singularity of the signal as the first type of defect feature; calculate the spatial gradient tensor of the background magnetic field distribution data and extract the maximum eigenvalue as the second type of defect feature, and at the same time, extract the extreme value of the local surface curvature as the third type of defect feature.

[0056] Specifically, for the 3D point cloud data of each detection point, a quadratic surface equation is fitted using the point set in the neighborhood of the detection point. The principal curvatures are derived by calculating the coefficients of the quadratic surface equation, and the average of the two principal curvatures is taken as the local surface curvature of the detection point. Then, a preset mapping rule is established: for example, when the local surface curvature is less than a threshold A, the number of modes K in the variational mode decomposition is set to 3, and a penalty factor is applied. Set to 2000; when the local surface curvature is greater than or equal to the threshold A, it indicates drastic geometric changes. In this case, the number of modes K is increased to 5, and the penalty factor is increased. Increase to 3500. Finally, use the parameter K, determined based on the local surface curvature of the detection point, and The variational mode decomposition algorithm is performed on the response impedance signal segment corresponding to the detection point to obtain a set of intrinsic mode functions, such as Figure 3 As shown.

[0057] For each decomposed intrinsic mode function, the optimal time delay parameter is determined by calculating the average mutual information of the signal. For example, the time corresponding to the first minimum point of the mutual information function can be taken as the optimal time delay parameter. The minimum embedding dimension *m* is determined using the pseudo-nearest neighbor method; for example, *m* can be set to 3. Using the determined... The intrinsic mode functions are reconstructed in time series with m to generate a point cloud in three-dimensional phase space, which is the attractor. A Vietoris-Rips complex is then constructed from the point cloud by gradually increasing the neighborhood radius. Calculate and record the number of one-dimensional annular holes in the point cloud topology at different radii, i.e., the first Betti number. The final result varies with radius. changing A sequence is a Betti number sequence that represents the inherent topological structure of a signal, and the Betti number sequence is used as a first-class defect feature.

[0058] Using the three-component magnetic field data obtained by a Hall sensor array in a unified surface coordinate system, the spatial partial derivatives of each magnetic field component along the U, V, and W coordinate axes are approximated by the difference in readings from adjacent sensors. These nine partial derivatives are then constructed into a third-order spatial gradient tensor matrix. Subsequently, eigenvalue decomposition is performed on the third-order spatial gradient tensor matrix, and the eigenvalue with the largest absolute value is extracted as the second type of defect feature. Simultaneously, for the local surface curvature sequence calculated from the three-dimensional geometric data along the contact line, a sliding window is used to find the local maxima, and these maxima are used as the third type of defect feature, such as... Figure 4 As shown.

[0059] In an optional embodiment, in S3, when performing variational mode decomposition on the response impedance signal, the absolute value of the local surface curvature is less than 0.01. At that time, the number of modes K in the variational mode decomposition is set to 4, and the penalty factor is... Set to 2000;

[0060] When the absolute value of the local surface curvature is between 0.01 Up to 0.05 When the interval is between, the number of modes K in the variational mode decomposition is set to 6, and the penalty factor is... Set to 1500;

[0061] When the absolute value of the local surface curvature is greater than 0.05 At that time, the number of modes K in the variational mode decomposition is set to 8, and the penalty factor is... Set to 1000.

[0062] Based on the three-dimensional geometric data obtained by laser scanning, the local curvature of the contact line surface corresponding to the current eddy current probe detection position is calculated in real time. This local curvature value reflects the smoothness or curvature of the contact line surface.

[0063] Subsequently, based on the preset mapping rules, the number of modes K and the penalty factor of the variational mode decomposition algorithm are adaptively selected according to the absolute magnitude of the local curvature values. These two key parameters:

[0064] If the absolute value of the local curvature is less than 0.01 This indicates that the surface is nearly flat, in which case a smaller number of modes K=4 and a higher penalty factor are used. =2000 is used to process signals.

[0065] If the absolute value of the local curvature is between 0.01 Up to 0.05 The value between these values ​​indicates a moderate degree of surface curvature; therefore, the parameters are adjusted to the number of modes K=6 and the penalty factor. =1500.

[0066] If the absolute value of the local curvature exceeds 0.05 This indicates a drastic change in surface geometry, thus requiring a higher modal number (K=8) and a lower penalty factor. =1000, to better capture more complex signal components.

[0067] Once the parameters are selected, use these parameters to perform variational mode decomposition on the response impedance signal at the current location.

[0068] In an optional embodiment, in S3, when extracting the Betti number sequence representing the singularity of the signal, the time delay parameters of each intrinsic mode function are calculated using the average mutual information method. Set the embedding dimension m to 5;

[0069] Based on time delay parameters Given an embedding dimension m, perform delayed embedding on the eigenmode functions to construct an attractor point cloud;

[0070] The 0th and 1st order continuous cohomology of the attractor point cloud were calculated based on the Vietoris-Rips complex, and the corresponding Betti numbers were extracted. and Sequence is used as a first-class defect feature.

[0071] The first step is to determine the key parameters for phase space reconstruction: For each eigenmode function time series obtained from variational mode decomposition, the average mutual information method is used to analyze its internal time correlation and calculate the optimal time delay parameters. At the same time, the embedding dimension m is directly set to 5. These two parameters... Together with m, it determines how to expand a one-dimensional time series into a higher-dimensional space.

[0072] The second step is to construct the attractor and perform continuous cohomology analysis: using the time delay parameter determined in the previous step. With an embedding dimension of m=5, delayed embedding is performed on the eigenmode functions to transform the data into a point cloud set in five-dimensional space; this point cloud set is the attractor. A continuous cohomology calculation method is applied to this attractor point cloud, specifically by constructing a Vietoris-Rips complex and analyzing its topological structure at different scales to calculate the 0th and 1st order continuous cohomology information. Two time series are extracted from the calculation results: the 0th order Betti number, describing the change in the number of connected components. Sequence and first-order Betty number describing the variation in the number of annular holes The two Betty number sequences were used as the first type of defect features.

[0073] In an optional embodiment, in S3, when calculating the spatial gradient tensor of the background magnetic field distribution data and extracting the maximum eigenvalue, a Hall sensor arranged in a 3×3 array centered on the eddy current probe is used to simultaneously acquire triaxial background magnetic field vector data at nine spatial locations. ;

[0074] Using the central difference method, the spatial partial derivatives of the magnetic field strength in three orthogonal directions are calculated based on the triaxial background magnetic field vector data of nine spatial locations, forming the gradient tensor matrix. :

[0075] ;

[0076] For the gradient tensor matrix Perform eigenvalue decomposition and extract the eigenvalue with the largest absolute value as the second type of defect feature.

[0077] Acquisition of spatial magnetic field data: A 3×3 array of nine triaxial Hall sensors is used, centered on an eddy current probe, to simultaneously measure the background magnetic field strength at nine discrete spatial points. Each triaxial Hall sensor outputs a value containing... and The magnetic field vector consists of three components, thus obtaining snapshot data of the magnetic field distribution within a local area.

[0078] Calculate the gradient tensor and extract features: Using the magnetic field vector data at these nine points, the central difference method is employed to estimate the spatial rate of change of the magnetic field in the U, V, and W directions, i.e., the partial derivatives of each magnetic field component with respect to each coordinate axis. For example, the change of the magnetic field along a certain direction is calculated using the magnetic field readings at the center point and adjacent points. All nine calculated partial derivative values ​​are organized into a 3x3 gradient tensor matrix G. Subsequently, eigenvalue decomposition is performed on this gradient tensor matrix G to obtain three eigenvalues. The eigenvalue with the largest absolute value is selected as the second type of defect feature describing the degree of background magnetic field distortion at the point.

[0079] S4. Construct a probabilistic graphical model with the first, second, and third types of defect features as input and the defect type and severity level as output. Perform probabilistic inference through the probabilistic graphical model, calculate the posterior probability of defects at each detection location, and identify the detection locations with posterior probability of defects exceeding a preset threshold as defect points.

[0080] Specifically, a Bayesian network is first constructed as a probabilistic graphical model. In this network, the first, second, and third types of defect features are used as parent nodes, while a discrete variable representing the defect state (such as no defect, wear, or crack) is set as a child node. During training, a database containing a large number of known defect samples is used to train the Bayesian network, and the conditional probability table in the network is learned through maximum likelihood estimation. In actual detection, the three types of feature values ​​extracted in real time are input into the trained Bayesian network, and inference algorithms such as belief propagation are used to calculate the posterior probability of the current detection location belonging to each state, such as no defect, wear, or crack. Finally, the posterior probabilities of all defect states are summed to obtain the total defect posterior probability. If the total probability value exceeds a preset confidence threshold (such as 0.9), the detection location is determined to be a defect point, and the specific type of defect can be determined based on the state with the highest probability.

[0081] In an optional embodiment, in S4, when constructing the probabilistic graphical model, a linear chain conditional random field is used as the probabilistic graphical model;

[0082] The continuous measurement points along the contact line detection path are regarded as a sequence. The first, second and third types of defect features corresponding to each measurement point in the sequence together constitute the observation value of the probabilistic graphical model.

[0083] The defect type is set as three discrete states: "surface crack", "internal crack" and "severe wear", and the severity level is set as three discrete states: "slight", "moderate" and "severe". These nine combined states are used as the output labels of the probabilistic graphical model.

[0084] The probabilistic graphical model is trained using a dataset of known defective samples.

[0085] For example, the dataset contains 500 known defect samples.

[0086] The first step is to obtain the model's structure and input / output: A linear-chain conditional random field is chosen as the probabilistic graphical model for the sequential defect classification task. The data from a series of continuous measurement points obtained during contact line inspection are considered as a sequence. For each measurement point in the sequence, the extracted first, second, and third type defect features are packaged into a feature vector, which serves as the model's observation or input. The model's output objective is to predict the defect state at each point, which is determined by both the defect type and severity level. The defect types are categorized into surface cracks, internal cracks, and severe wear; the severity levels are categorized into minor, moderate, and severe, resulting in a total of nine possible combined labels.

[0087] The second step is model training: Prepare a training dataset containing 500 known defect samples. Each sample includes input features of a detection sequence and a corresponding correct defect label sequence pre-annotated by experts. Input this data into a linear chain conditional random field model, and use a learning algorithm to adjust the model's internal parameters so that it can best learn the mapping relationship between the input feature sequence and the output label sequence. After training, the model can be used to automatically determine the defect type and severity level of new, unknown detection data sequences.

[0088] In addition, in the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

Claims

1. A method for detecting defects in a railway contact wire based on eddy currents, characterized in that, The method comprises the following steps: S1, a preset frequency-sweep chirp baseband signal is phase-modulated by a pseudo-random binary sequence to generate a composite excitation signal, and the composite excitation signal is used to drive an eddy current probe; S2, the response impedance signal of the eddy current probe, the three-dimensional geometric data of the contact line and the circumferential background magnetic field distribution data of the eddy current probe are synchronously collected; based on the three-dimensional geometric data, a transient surface coordinate system of the contact line is constructed, and the response impedance signal and the background magnetic field distribution data are mapped to the transient surface coordinate system to obtain multi-modal sensing data unified in time and space; S3, the local surface curvature of the three-dimensional geometric data is calculated, the parameters of variational modal decomposition are adaptively adjusted according to the local surface curvature, the variational modal decomposition is performed on the response impedance signal to obtain a group of eigenmode functions; the phase space reconstruction is performed on each eigenmode function to construct an attractor, the persistent homology of the attractor is calculated, and a Betti number sequence representing signal singularity is extracted as a first type of defect feature; The spatial gradient tensor of the background magnetic field distribution data is calculated, and the maximum eigenvalue is extracted as a second type of defect feature; meanwhile, the extreme value of the local surface curvature is extracted as a third type of defect feature; S4, a probabilistic graph model is constructed with the first, second and third types of defect features as inputs and the defect type and severity level as outputs, probability inference is performed through the probabilistic graph model, the defect posterior probability of each detection position is calculated, and the detection position with a defect posterior probability higher than a preset threshold is recognized as a defect point; In S3, when the absolute value of the local surface curvature is less than 0.01 when performing the variational mode decomposition on the response impedance signal, the mode number K of the variational mode decomposition is set to 4, and the penalty factor is set to 2000. When the absolute value of the local surface curvature is between 0.01 and 0.05 , the number of modes K of the variational mode decomposition is set to 6, and the penalty factor is set to 1500. When the absolute value of the local surface curvature is greater than 0.05 The number of modes K of the variational mode decomposition is set to 8, and the penalty factor is set to 1000.

2. The eddy current based railway contact line defect detection method according to claim 1, characterized in that, In S1, when the composite excitation signal is generated, an m sequence is generated by using a linear feedback shift register with a specific order to generate an m sequence, and the m sequence is used as the pseudo-random binary sequence; The starting frequency, the terminal frequency and the sweep period of the frequency-sweep chirp baseband signal are obtained; The m sequence is used to phase-modulate the frequency-sweep chirp baseband signal in a binary phase shift keying mode to obtain the composite excitation signal.

3. The eddy current based railway contact line defect detection method according to claim 2, characterized in that, The linear feedback shift register is a 10-order linear feedback shift register.

4. The eddy current based railway contact line defect detection method according to claim 2, characterized in that, The starting frequency of the frequency-sweep chirp baseband signal is 5 kHz, the terminal frequency is 50 kHz, and the sweep period is 1 ms.

5. The eddy current based railway contact line defect detection method according to claim 1, characterized in that, In S2, when the response impedance signal and the background magnetic field distribution data are mapped to the transient surface coordinate system, at each sampling time, a center line tangent vector of the contact line laser scanning profile is extracted as a U-axis, a normal vector of the contact line profile wear surface is extracted as a V-axis, a W-axis is obtained through the cross product of the U-axis and the V-axis, and a right-hand transient surface coordinate system is constructed through the U-axis, the V-axis and the W-axis; At the corresponding sampling time, the response impedance signal of the eddy current probe is mapped to the origin of the right-hand transient surface coordinate system, and the background magnetic field distribution data is mapped to the right-hand transient surface coordinate system according to the sampling position.

6. The eddy current based railway contact line defect detection method according to claim 1, characterized in that, In S3, when extracting the BEM sequence representing the singularity of the signal, the average mutual information method is used to calculate the time delay parameter of each intrinsic mode function , the embedding dimension m is set to 5; According to the time delay parameter And the embedding dimension m, the delay embedding is performed on the intrinsic mode function to construct an attractor point cloud; The 0th and 1st persistent homology of the attractor point cloud is calculated based on Vietoris-Rips complex, and the corresponding Betti numbers are extracted and the sequence as the first type of defect feature.

7. The eddy current based railway contact line defect detection method according to claim 6, characterized in that, In S3, when calculating the spatial gradient tensor of the background magnetic field distribution data and extracting the maximum eigenvalue, a 3x3 array of Hall sensors centered on the eddy current probe is used to synchronously collect three-axis background magnetic field vector data at nine spatial position points ; The spatial partial derivatives of the magnetic field strength in three orthogonal directions are calculated from the three-axis background magnetic field vector data of nine spatial position points by using the central difference method to form a gradient tensor matrix : ； performing eigenvalue decomposition on the gradient tensor matrix extracting the eigenvalue with the largest absolute value as the second type of defect feature.

8. The eddy current based railway contact line defect detection method according to any one of claims 1-7, characterized in that, In S4, when the probabilistic graph model is constructed, a linear chain conditional random field is used as the probabilistic graph model; A sequence of continuous measurement points along the contact line detection path is regarded as an observation value of the probabilistic graph model, and the first, second and third types of defect features corresponding to each measurement point in the sequence jointly constitute the observation value. The defect type is set as three discrete states of "surface crack", "internal crack" and "severe wear", the severity level is set as three discrete states of "slight", "moderate" and "severe", and the nine combined states are used as output labels of the probabilistic graph model; The probabilistic graph model is parameter trained using a data set of known defect samples.

9. The eddy current based railway contact line defect detection method according to claim 8, characterized in that, The number of the known defect samples is 500.

Images