A system and method for evaluating magnetic coercivity damage of a metal member

Abstract

The application relates to the technical field of metal component nondestructive testing, and discloses a metal component magnetic coercivity damage evaluation system and method, wherein the metal component magnetic coercivity damage evaluation method comprises the following steps: performing repeated standardization magnetization cycle collection and synchronous average processing on each measuring point to generate high signal-to-noise ratio average magnetic hysteresis loop data; extracting high-order statistical features of a local interval of the magnetic hysteresis loop to generate a local skewness value sequence, a local kurtosis value sequence and a cycle interval micro-variability value sequence; generating a multi-channel high-order statistical feature space field through Kriging interpolation; performing time sequence difference and gradient vector field calculation on the multi-time and multi-channel high-order statistical feature space field; performing DBSCAN density clustering on the cross-channel fusion gradient vector to identify an early damage active evolution front and extract front features; and outputting a metal component magnetic coercivity damage evaluation report.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing and service safety assessment of metal components, and more specifically, to a magnetic coercive damage assessment system and method for metal components. Background Technology

[0002] In the field of nondestructive testing and service safety assessment of metal components, magnetic coercivity is an important magnetic parameter reflecting changes in the microstructure of ferromagnetic metallic materials. Existing methods for assessing magnetic coercivity damage in metal components typically involve performing standardized magnetization cycles at various measurement points on the component surface, collecting hysteresis loop data, extracting coercivity values, comparing the coercivity values ​​at each point with standard values ​​under known material conditions, and determining the degree of local damage based on the deviation. For large, safety-critical metal components requiring multiple periodic inspections, the spatial distribution field of coercivity at each inspection time is further subdivided pixel by pixel to calculate the spatial gradient vector field, which tracks the expansion direction and rate of the damaged area, thereby assessing the service safety of the component.

[0003] However, when damage to metallic components is in its very early stages, i.e., when dislocation substructures are just beginning to form, the change in the absolute value of coercivity is only one to three percent of the initial value. This minute change in the absolute value of coercivity is easily masked by measurement noise and environmental temperature fluctuations, resulting in the inability to distinguish between true damage change signals and spurious change signals in the spatial field difference results based on the absolute value of coercivity, rendering spatial evolution tracking completely ineffective. Therefore, existing magnetic coercivity damage assessment methods for metallic components suffer from insufficient detection sensitivity and lack of spatial localization in the early microscopic damage stage, failing to simultaneously meet the assessment requirements of high-sensitivity early damage detection and damage spatial evolution tracking. Summary of the Invention

[0004] This invention provides a magnetic coercive damage assessment system and method for metal components, which solves the technical problems in related technologies such as the difficulty in accurately locating early damage to metal components and the difficulty in effectively identifying and warning of the damage evolution front.

[0005] This invention discloses a method for assessing magnetic coercive damage to metal components, comprising:

[0006] Multiple standardized magnetization cycles are performed on each measurement point on the surface of the metal component according to a preset spatial grid to acquire multiple hysteresis loop raw data of each measurement point. The multiple hysteresis loop raw data of each measurement point are synchronously averaged to generate high signal-to-noise ratio average hysteresis loop data of each measurement point.

[0007] For the rising branch data of the high signal-to-noise ratio average hysteresis loop at each measurement point, multiple local interval windows are uniformly divided on the magnetic field strength axis. The local skewness value and local kurtosis value are calculated for the magnetization intensity data in each local interval window. Based on the original data of multiple hysteresis loops at each measurement point, the inter-cycle microvariability value is calculated in each local interval window, generating the local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariability value sequence for each measurement point.

[0008] The local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariation value sequence of each measurement point are arranged according to the two-dimensional spatial coordinates of each measurement point on the surface of the component. Kriging interpolation is performed on each channel to generate a multi-channel high-order statistical feature space field at the current detection time.

[0009] The system obtains the multi-channel high-order statistical feature spatial field of historical detection time, calculates the pixel-by-pixel difference of the high-order statistical feature spatial field of each channel between the current detection time and the adjacent historical detection time to obtain the high-order statistical feature change field of each channel, calculates the spatial gradient vector of each pixel position in the high-order statistical feature change field of each channel, and generates the high-order statistical feature change gradient vector field of each channel.

[0010] All channel gradient vectors at each pixel location are concatenated into a fused gradient vector. Density clustering is performed on the fused gradient vectors at each pixel location on the component surface to identify spatially continuous regions with consistent gradient directions and significant amplitudes, which are then marked as active evolution fronts of early damage. Front features of each active evolution front of early damage are extracted to generate a feature set of spatial evolution fronts of early damage in metal components.

[0011] A magnetic coercive damage assessment report for metal components is generated based on the feature set of early damage spatial evolution front of the metal components.

[0012] Furthermore, the synchronous averaging process refers to taking the arithmetic mean of the magnetization values ​​of multiple hysteresis loops at the same measurement point at the same magnetic field strength sampling point; the standardized magnetization cycle refers to applying a periodically changing magnetic field with a preset waveform to the measurement point, so that the material undergoes a complete magnetization process from a negative saturation magnetization state to a positive saturation magnetization state and then back to a negative saturation magnetization state. The standardized magnetization cycle of each measurement point adopts the same magnetic field strength variation range, the same magnetic field variation rate, and the same sampling time interval.

[0013] Furthermore, the local skewness value is calculated as follows: for the average magnetization value corresponding to each magnetic field strength sampling point within the local interval window, the ratio of the cube of the magnetization deviation to the cube of the magnetization standard deviation is calculated; the local kurtosis value is calculated as follows: for the average magnetization value corresponding to each magnetic field strength sampling point within the local interval window, the ratio of the fourth power of the magnetization deviation to the fourth power of the magnetization standard deviation is calculated; the inter-cycle microvariability value is calculated as follows: for multiple hysteresis loops at the same measurement point, the mean magnetization value of each hysteresis loop within the local interval window is calculated, and then the variance between these multiple mean magnetization values ​​is calculated.

[0014] Furthermore, before performing Kriging interpolation on each channel, Z-score standardization is performed on the local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariance value sequence of each measurement point, so that the feature values ​​of each channel are unified into dimensionless values ​​with a mean of zero and a standard deviation of one. Each channel in the multi-channel high-order statistical feature space field corresponds to a statistical feature type under a local interval window number, and the total number of channels is three times the number of local interval windows.

[0015] Furthermore, before calculating the pixel-by-pixel difference, spatial registration processing is performed on the multi-channel high-order statistical feature space field at all detection times. The spatial registration processing includes: using the fixed marker points preset on the component surface as control points, selecting at least three non-collinear fixed marker points, identifying the actual pixel coordinates of the control points in the multi-channel high-order statistical feature space field at each time, establishing a correspondence with the reference coordinates of the corresponding control points in a unified reference coordinate system, determining the affine transformation matrix by solving a system of linear equations, applying the affine transformation matrix to all pixel positions in the multi-channel high-order statistical feature space field at each time for geometric alignment, and obtaining the feature values ​​of the mapped pixel positions through bilinear interpolation to generate the registered multi-channel high-order statistical feature space field.

[0016] Furthermore, after spatial registration and before pixel-by-pixel differencing, environmental drift correction is performed on the high-order statistical feature spatial field of each channel at each time. The environmental drift correction includes: calculating the regional mean of the high-order statistical feature spatial field of each channel at each time within a reference area where the component is known to be undamaged; using the difference between the regional mean of each channel at the current time and the reference time as the systematic environmental drift estimate of that channel; and subtracting the systematic environmental drift estimate of the corresponding channel at the corresponding time from the feature value of all pixel positions in the high-order statistical feature spatial field of each channel at each time to generate the environmentally corrected multi-time multi-channel high-order statistical feature spatial field.

[0017] Furthermore, the density clustering employs the DBSCAN algorithm. During the clustering process, the weighted sum of the Euclidean distance between the fused gradient vectors and the spatial distance between the pixel spatial coordinates is used as the joint similarity metric distance. The joint similarity metric distance is the sum of the spatial Euclidean distance and the fused gradient vector Euclidean distance multiplied by their respective weight coefficients, where the sum of the spatial distance weight coefficient and the gradient distance weight coefficient is one. When the joint similarity metric distance between two pixel positions does not exceed a preset neighborhood radius, the two pixel positions are considered neighbors. When the number of pixels contained in the neighborhood of a pixel position is not less than a preset minimum number of neighborhood points, the pixel position is determined to be a core point. All pixel positions that are density-reachable from the core point together constitute a cluster, and each cluster that satisfies the minimum density condition is marked as an early damage active evolution front.

[0018] Furthermore, the leading edge features include the main expansion direction angle, the average expansion rate, and the leading edge length; the main expansion direction angle is the direction angle of the weighted composite vector obtained by summing the fused gradient vectors of all pixel positions within the active evolution leading edge of early damage in magnitude weight; the average expansion rate is the arithmetic mean of the magnitudes of the fused gradient vectors of all pixel positions within the active evolution leading edge of early damage; and the leading edge length is the difference between the maximum and minimum values ​​of the projection values ​​of the coordinates of all pixel positions within the active evolution leading edge of early damage onto the unit vector of the main expansion direction.

[0019] Furthermore, based on obtaining the feature set of early damage spatial evolution fronts of metal components for multiple detection cycles, the method further includes: performing linear regression trend extrapolation on the main expansion direction angle and average expansion rate of each early damage active evolution front in multiple detection cycles to predict the expected spatial location and expected coverage of each early damage active evolution front in the next detection cycle; superimposing and comparing the predicted location of each early damage active evolution front with the spatial coordinates of the critical load-bearing area of ​​the component to calculate the expected time for each early damage active evolution front to reach the nearest critical load-bearing area boundary at the current expansion rate; marking early damage active evolution fronts with expected arrival times lower than a preset safe time threshold as high-risk warning objects, and outputting corresponding graded warning information in the magnetic coercive damage assessment report of the metal component, wherein the graded warning information includes the risk level of each early damage active evolution front, the expected time to reach the critical load-bearing area, and the corresponding critical load-bearing area identifier.

[0020] This invention provides a magnetic coercive damage assessment system for metal components, comprising:

[0021] The hysteresis loop acquisition and averaging module is used to perform multiple repeated standardized magnetization cycles to acquire raw data of multiple hysteresis loops at each measurement point on the surface of the metal component, and to perform synchronous averaging of the raw data of multiple hysteresis loops at each measurement point to generate high signal-to-noise ratio averaged hysteresis loop data for each measurement point.

[0022] The advanced statistical feature extraction module is used to uniformly divide the rising branch data of the high signal-to-noise ratio average hysteresis loop at each measurement point into multiple local interval windows on the magnetic field strength axis. It calculates the local skewness value and local kurtosis value for the magnetization intensity data in each local interval window, and calculates the inter-cycle microvariability value in each local interval window based on the original data of multiple hysteresis loops at each measurement point, generating the local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariability value sequence for each measurement point.

[0023] The spatial field generation module is used to arrange the local skewness value sequence, local kurtosis value sequence and inter-cycle microvariation value sequence of each measurement point according to the two-dimensional spatial coordinates of each measurement point on the surface of the component, and perform Kriging interpolation on each channel to generate a multi-channel high-order statistical feature spatial field at the current detection time.

[0024] The temporal difference and gradient calculation module is used to obtain the multi-channel high-order statistical feature spatial field of historical detection time. It calculates pixel-by-pixel difference of the high-order statistical feature spatial field of each channel between the current detection time and the adjacent historical detection time to obtain the high-order statistical feature change field of each channel. It also calculates the spatial gradient vector of each pixel position in the high-order statistical feature change field of each channel to generate the high-order statistical feature change gradient vector field of each channel.

[0025] The density clustering and front identification module is used to concatenate all channel gradient vectors at each pixel position into a fused gradient vector, perform density clustering on the fused gradient vectors at each pixel position on the component surface, identify spatially continuous regions with consistent gradient directions and significant amplitudes and mark them as active evolution fronts of early damage, extract the front features of each active evolution front of early damage, and generate a feature set of spatial evolution fronts of early damage in metal components.

[0026] The assessment report generation module is used to generate a magnetic coercive damage assessment report for the metal component based on the feature set of the early damage spatial evolution front of the metal component.

[0027] The beneficial effects of this invention are as follows:

[0028] This invention replaces the absolute value of coercivity with the higher-order statistical morphological features of local intervals of hysteresis loops as the basic physical quantity for damage assessment. After extending the higher-order statistical features to a multi-channel spatial field, it performs temporal difference, gradient calculation, and cross-channel density clustering. This solves the technical problems of insufficient detection sensitivity and inability to distinguish between real damage signals and pseudo-change signals in existing methods in the early stage of micro-damage. It achieves the technical effect of being able to generate an effective numerical response to the small morphological distortion of hysteresis loops that is masked by measurement noise, and being able to separate and identify the true damage evolution front from the noise background. Attached Figure Description

[0029] Figure 1 This is a flowchart of the magnetic coercive damage assessment method for metal components provided in the embodiments of the present invention;

[0030] Figure 2 This is a schematic diagram of the magnetization intensity distribution at measurement point p47 at H201 after 25 cycles, provided in an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram illustrating the synchronous average noise reduction effect provided by an embodiment of the present invention: the change of noise standard deviation with the number of cycles;

[0032] Figure 4 This is a schematic diagram comparing the standardized values ​​of higher-order statistical features of each window at measurement point p47 provided in an embodiment of the present invention;

[0033] Figure 5 This is a schematic diagram of the spatial distribution (Kriging interpolation control points) of the third window skewness channel in the weld heat-affected zone provided in an embodiment of the present invention;

[0034] Figure 6 This is a schematic diagram of the temporal difference distribution of the third window skewness channel of the p47 neighboring pixels provided in an embodiment of the present invention;

[0035] Figure 7 This is a schematic diagram illustrating the multi-cycle evolution trend of the damage front F-01 provided in an embodiment of the present invention;

[0036] Figure 8 This is a schematic diagram illustrating the evolution trend of the main expansion direction angle of the damage front F-01 provided in an embodiment of the present invention. Detailed Implementation

[0037] In the field of nondestructive testing and service safety assessment of metal components, magnetic coercivity is an important magnetic parameter reflecting changes in the microstructure of ferromagnetic metallic materials. Existing methods for assessing magnetic coercivity damage in metal components typically involve performing standardized magnetization cycles at various measurement points on the component surface, collecting hysteresis loop data, extracting coercivity values, comparing the coercivity values ​​at each point with standard values ​​under known material conditions, and determining the degree of local damage based on the deviation. For large, safety-critical metal components requiring multiple periodic inspections, the spatial distribution field of coercivity at each inspection time is further subdivided pixel by pixel to calculate the spatial gradient vector field, which tracks the expansion direction and rate of the damaged area, thereby assessing the service safety of the component.

[0038] However, when damage to metallic components is in its very early stages, i.e., when dislocation substructures are just beginning to form, the change in the absolute value of coercivity is only one to three percent of the initial value. This minute change in the absolute value of coercivity is easily masked by measurement noise and environmental temperature fluctuations, resulting in the inability to distinguish between true damage change signals and spurious change signals in the spatial field difference results based on the absolute value of coercivity, rendering spatial evolution tracking completely ineffective. Therefore, existing magnetic coercivity damage assessment methods for metallic components suffer from insufficient detection sensitivity and lack of spatial localization in the early microscopic damage stage, failing to simultaneously meet the assessment requirements of high-sensitivity early damage detection and damage spatial evolution tracking.

[0039] According to an embodiment of this invention, a method for assessing magnetic coercive damage to metal components is provided. It should be understood that the execution entity of the method is a computer system with data acquisition and processing capabilities, and the computer system is communicatively connected to a magnetization acquisition device. The magnetization acquisition device includes a magnetization coil capable of applying a standardized magnetization cycle to each measurement point on the surface of the metal component, and a magnetic sensor for acquiring hysteresis loop data at each measurement point. Several fixed marker points are pre-set on the surface of the component for spatial positioning reference between each detection. At least one embodiment of this invention discloses a method for assessing magnetic coercive damage to metal components, such as... Figures 1-8 As shown, it includes the following steps:

[0040] Step 1: Collect hysteresis loop data at various measurement points on the surface of the metal component, and generate high signal-to-noise ratio average hysteresis loop data through synchronous averaging processing;

[0041] For each measurement point on the surface of the metal component, divided according to a preset spatial grid, a magnetization acquisition device performs N repeated standardized magnetization cycles to acquire raw data of N hysteresis loops for each measurement point. The raw data of the N hysteresis loops for each measurement point are then synchronously averaged to generate high signal-to-noise ratio averaged hysteresis loop data for each measurement point.

[0042] Synchronous averaging refers to taking the arithmetic mean of the magnetization values ​​of N hysteresis loops at the same measurement point at the same magnetic field strength sampling point. Let the p-th measurement point be at magnetic field strength sampling point H. j The magnetization value at the nth cycle is M. p , j , n Then the measurement point is at H j The average magnetization at that location is:

[0043] ;

[0044] Where p is the measurement point number, j is the magnetic field strength sampling point number, n is the cycle number, and N is the total number of repeated magnetization cycles. Through synchronous averaging, the standard deviation of random measurement noise is reduced to [a fraction of] the original single-measurement noise standard deviation. This increases the signal-to-noise ratio by several times, thereby improving the signal-to-noise ratio of high signal-to-noise ratio average hysteresis loop data.

[0045] Furthermore, standardized magnetization cycling refers to applying a periodically changing magnetic field of a preset waveform to the measurement point, causing the material to undergo a complete magnetization process from a negative saturation magnetization state to a positive saturation magnetization state and then back to a negative saturation magnetization state. Specifically, the magnetic field strength generated by the magnetization coil at the measurement point gradually increases from a negative saturation value to a positive saturation value according to a preset time function, and then gradually decreases from a positive saturation value to a negative saturation value, completing one complete magnetization cycle. The magnetic sensor synchronously collects the magnetization intensity response of the material during the magnetization cycle, forming the correspondence data between magnetic field strength and magnetization intensity, i.e., the raw hysteresis loop data. The standardized magnetization cycle at each measurement point uses the same magnetic field strength variation range, the same magnetic field variation rate, and the same sampling time interval to ensure the comparability of hysteresis loop data between different measurement points.

[0046] It should be noted that the number of repetitive magnetization cycles N is a preset positive integer. The value of the number of cycles N is determined based on the target signal-to-noise ratio improvement factor, and the typical range of the number of cycles N is 10 to 50.

[0047] Step 2: Extract the local interval high-order statistical features of the high signal-to-noise ratio average hysteresis loop at each measurement point to generate local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariation value sequence;

[0048] For the rising branch data of the high signal-to-noise ratio average hysteresis loop at each measurement point, w local interval windows are uniformly divided along the magnetic field strength axis. Local skewness and local kurtosis values ​​are calculated for the magnetization data within each local interval window. Simultaneously, based on the original data of N repeating hysteresis loops at each measurement point, the inter-cycle microvariability values ​​within each local interval window are calculated. A sequence of local skewness values ​​{S} distributed along the magnetic field strength axis at each measurement point is generated. p ,1,Sp ,2,…,S p , w}, Local kurtosis sequence {K p ,1,K p ,2,...,K p , w} and the sequence of inter-cycle microvariability values ​​{V p ,1,V p ,2,…,V p , w}. Among them, S p 1, S p 2, S p , w Let K be the local skewness value of the p-th measurement point in the 1st, 2nd, and wth local interval windows, respectively. p 1, K p 2, K p , w V represents the local kurtosis values ​​of the p-th measurement point in the 1st, 2nd, and wth local interval windows, respectively. p 1, V p 2, V p , w These represent the inter-cycle microvariation values ​​of the p-th measurement point in the 1st, 2nd, and wth local interval windows, respectively.

[0049] Specifically, for the w-th local interval window of the p-th measurement point, let the local interval window contain Q magnetic field strength sampling points, and the corresponding average magnetization value is ,in , , These represent the average magnetization values ​​corresponding to the 1st, 2nd, and Qth magnetic field strength sampling points within the w-th local interval window of the p-th measurement point. The mean magnetization value within the local interval window is μ. p , w The standard deviation of magnetization within the local interval window is σ. p , w The definitions of each quantity are as follows:

[0050] ;

[0051] ;

[0052] Where i is the number of the magnetic field intensity sampling point within the local interval window, and Q is the total number of magnetic field intensity sampling points contained within the local interval window.

[0053] The formula for calculating the local skewness value is:

[0054] ;

[0055] The formula for calculating local kurtosis is:

[0056] ;

[0057] Among them, S p , w K represents the local skewness value of the w-th local interval window at the p-th measurement point. p , w Let w be the local kurtosis value of the w-th local interval window at the p-th measurement point, where w is the local interval window number.

[0058] It should be noted that the aforementioned local skewness value S p , w and local kurtosis value K p , w In the calculation formula, the numerator is the cube or fourth power of the magnetization deviation, and the denominator is the standard deviation of magnetization σ. p , w The corresponding powers of the two have the same dimensions, and the local skewness value S is obtained by dividing them. p , w and local kurtosis value K p , w All are dimensionless pure numerical values, so there is no issue of dimension mismatch.

[0059] Furthermore, the local skewness value reflects the degree of asymmetry in the magnetization intensity distribution within a local interval window relative to the mean. When the material microstructure is uniform, the flipping behavior of magnetic domains within this magnetic field intensity range is statistically symmetrical, the magnetization intensity distribution is nearly symmetrical, and the local skewness value is close to zero. When early damage leads to an increase in local dislocation density, the pinning strength of some magnetic domain walls changes, causing the magnetization response within this interval to shift towards higher or lower values, and the local skewness value deviates from zero. The direction and magnitude of the deviation in the local skewness value reflect the type and degree of damage.

[0060] Furthermore, the local kurtosis value reflects the sharpness and tail thickness of the magnetization intensity distribution within a local window. When the material microstructure is uniform, the statistical distribution of domain flipping behavior is close to a normal distribution, and the local kurtosis value is close to 3. When early damage introduces microstructure inhomogeneity, the flipping resistance of some domains deviates significantly from the average level, resulting in sharper peaks or thicker tails in the magnetization intensity distribution, and the local kurtosis value deviates from 3. The magnitude of the deviation in the local kurtosis value reflects the degree of microstructure inhomogeneity.

[0061] The calculation of inter-cycle microvariability values ​​is based on the raw data of N repetitive hysteresis loops at each measurement point. For the w-th local interval window at the p-th measurement point, the mean magnetization of each hysteresis loop within that local interval window is first calculated. (n = 1, 2, ..., N), where Let N be the mean magnetization intensity of the nth hysteresis loop within the wth local interval window at the pth measurement point, and then calculate the variance among the means of these N local interval windows:

[0062] ;

[0063] in, V is the arithmetic mean of the means of N local interval windows. p , w This represents the inter-cycle microvariability value of the w-th local interval window at the p-th measurement point.

[0064] It should be noted that the rising branch of the hysteresis loop mentioned above refers to the part of the hysteresis loop where the magnetic field strength monotonically increases from the negative saturation value to the positive saturation value.

[0065] It should be noted that the number W of the aforementioned local interval windows is a preset positive integer. Each local interval window is arranged with equal width on the magnetic field strength axis, and there is no overlap between adjacent local interval windows. The magnetic field strength range covered by each local interval window is equal to the span of the total ascending branch magnetic field strength. .

[0066] It should be noted that the above inter-cycle microvariability value V p , w Unlike the local skewness and local kurtosis of high signal-to-noise ratio average hysteresis loops, the intercycle microvariability value V p , w It directly measures the fluctuation amplitude of the hysteresis loop shape between each cycle during repeated magnetization at the same measurement point, reflecting the instability of the magnetic domain motion of the material within the range of magnetic field strength.

[0067] Furthermore, the intercycle microvariability reflects the repeatability of domain movement during repeated magnetization. When the material microstructure is stable, the domain flipping paths are highly consistent in each magnetization cycle, and the intercycle microvariability is small. When early damage introduces movable dislocations or microcracks, the pinning positions and flipping paths of the domain walls during repeated magnetization exhibit random fluctuations, increasing the intercycle microvariability. The magnitude of this increase reflects the activity of microdefects.

[0068] It should be noted that the inter-cycle microvariability value V p , w The dimension of is the square of the magnetization, while the local skewness value S p , w and local kurtosis value K p , wAll values ​​are dimensionless pure numerical values, and there are dimensional differences among the three types of feature quantities. To eliminate the impact of dimensional differences on subsequent multi-channel spatial field fusion and cross-channel clustering operations, before performing Kriging interpolation to generate multi-channel high-order statistical feature spatial fields in step 3, Z-score standardization is performed on the local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariance value sequence of each measurement point. This unifies the feature values ​​of each channel into dimensionless values ​​with a mean of zero and a standard deviation of one before proceeding with subsequent processing.

[0069] Step 3: Arrange the higher-order statistical feature values ​​of each measurement point according to spatial coordinates, and generate a multi-channel higher-order statistical feature spatial field by kriging interpolation;

[0070] The Z-score-normalized sequences of local skewness, local kurtosis, and inter-cycle microvariation values ​​for each measurement point are arranged according to their two-dimensional spatial coordinates on the component surface. Kriging interpolation is then performed on each channel to generate a continuous distribution field with uniform spatial resolution covering the component surface, thus obtaining the multi-channel high-order statistical feature spatial field at the current detection time.

[0071] The input to Kriging interpolation is the two-dimensional spatial coordinates of each measurement point and the standardized feature value of the corresponding channel of each measurement point. The output is the continuous interpolated feature value of each pixel position on the uniform grid covering the surface of the component.

[0072] Furthermore, Kriging interpolation is an optimal linear unbiased estimation method based on spatial autocorrelation. For the pixel position to be interpolated, Kriging interpolation first calculates the spatial autocorrelation function based on the spatial distribution of known measurement points, and then uses the spatial autocorrelation function to determine the weight coefficients of each known measurement point at the position to be interpolated, minimizing the variance of the interpolation estimate. Specifically, for the feature value of a certain channel at the pixel position to be interpolated, the Kriging interpolation estimate is a weighted linear combination of the feature values ​​of that channel at each measurement point. The weight coefficients are obtained by solving the Kriging equations, and the coefficient matrix of the Kriging equations consists of the spatial autocorrelation function values ​​between measurement points, with the constraint that the sum of the weight coefficients equals 1. Through Kriging interpolation, the feature values ​​of discrete measurement points are extended into a continuous spatial field covering the surface of the component, and the interpolation result is exactly equal to the measured value at the known measurement point position, and smoothly transitions between measurement points.

[0073] It should be noted that each channel in the above multi-channel higher-order statistical feature space field corresponds to a statistical feature type under a local interval window number w. The total number of channels is 3w, namely w local skewness value space fields, w local kurtosis value space fields, and w inter-cycle microvariance value space fields.

[0074] Step 4: Perform temporal difference and gradient vector field calculation on the spatial field of high-order statistical features of multiple time points and multiple channels to generate gradient vector fields of high-order statistical features of each channel;

[0075] Obtain multi-channel high-order statistical feature spatial fields generated using the same procedures as steps 1 to 3 at at least two historical detection times. For each channel's high-order statistical feature spatial field between the current detection time and adjacent historical detection times, calculate pixel-by-pixel differences to obtain the high-order statistical feature change field for each channel. Calculate the spatial gradient vector for each pixel position in the high-order statistical feature change field of each channel to generate the high-order statistical feature change gradient vector field for each channel.

[0076] Let the current detection time be t, and the adjacent historical detection time be t-1. The higher-order statistical feature change value of the c-th channel at pixel position (x, y) is:

[0077] ;

[0078] in, Let be the feature value of the c-th channel at pixel position (x, y) at time t. Let t-1 be the feature value of the c-th channel at pixel position (x, y), where c is the channel number, x and y are the horizontal and vertical coordinates of the pixel position, respectively, t and t-1 represent the time numbers of two adjacent detections, and the time interval between the two detections is the actual detection period.

[0079] Furthermore, the time difference ΔF c (x, y) reflects the change in the feature value of the c-th channel at pixel position (x, y) between two adjacent detection cycles. Since the multi-channel high-order statistical feature space field generated in step 3 is arranged in order of detection time number t in the time dimension, each time point corresponds to a complete multi-channel high-order statistical feature space field obtained from a single component surface detection. Therefore, the temporal difference ΔF c (x, y) directly reflects the temporal evolution of the channel's feature value within a detection period, and is the temporal dimension basis for subsequent gradient calculation and cluster analysis.

[0080] The spatial gradient vector of each channel at pixel position (x, y) is:

[0081] ;

[0082] in, and These are the higher-order statistical characteristic change fields ΔF c The partial derivatives in the x and y directions are obtained by approximating the difference between adjacent pixels.

[0083] Furthermore, the partial derivatives of the spatial gradient vector are calculated using the central difference scheme. For pixel position (x, y), the partial derivative of the spatial gradient vector in the x-direction is... The partial derivative of the spatial gradient vector in the y-direction is For pixel locations at the boundaries of the multi-channel higher-order statistical feature spatial field, partial derivatives are calculated using forward or backward differencing schemes. The direction of the spatial gradient vector points to the direction in which the feature value of that channel increases the most spatially, and the magnitude of the spatial gradient vector reflects the rate of change in that direction.

[0084] Furthermore, to eliminate the spatial offset introduced by the measurement grid alignment deviation caused by the repositioning of the detection device between detections, spatial registration processing is performed on the multi-channel high-order statistical feature spatial field at all detection times before calculating the pixel-by-pixel temporal difference. Specifically, using preset fixed marker points on the component surface as control points, affine transformation geometric alignment is performed on the multi-channel high-order statistical feature spatial field at each time. The input of the affine transformation is the correspondence between the actual pixel coordinates of the fixed marker points and the unified reference coordinates in the multi-channel high-order statistical feature spatial field at each time, and the output is the registered multi-channel high-order statistical feature spatial field after translation, rotation, and scaling correction. After spatial registration processing, the same pixel position in the multi-channel high-order statistical feature spatial field at each time corresponds to the same physical position on the component surface, and the subsequent pixel-by-pixel difference results can accurately reflect the temporal changes of feature values ​​at the same physical position.

[0085] Furthermore, affine transformation achieves geometric alignment by establishing a linear mapping between the pixel coordinate system of the multi-channel high-order statistical feature space field at each time step and a unified reference coordinate system. At least three non-collinear fixed marker points on the component surface are selected as control points. The actual pixel coordinates of these control points are identified in the multi-channel high-order statistical feature space field at each time step, and a correspondence is established with the reference coordinates of the corresponding control points in the unified reference coordinate system. The six parameters of the affine transformation matrix are determined by solving a system of linear equations. The affine transformation matrix includes a combination of translation, rotation, scaling, and shearing transformations. The affine transformation matrix is ​​applied to all pixel positions in the multi-channel high-order statistical feature space field at each time step, mapping the coordinates of each pixel position to the unified reference coordinate system. The eigenvalues ​​of the mapped pixel positions are obtained through bilinear interpolation, generating the registered multi-channel high-order statistical feature space field. The registered multi-channel high-order statistical feature space field at each time step achieves precise correspondence between the pixel coordinates and the reference coordinates at the control point positions, while minimizing and correcting geometric distortion at non-control point positions.

[0086] Furthermore, based on spatial registration, to further eliminate the global shift caused by systematic factors such as environmental temperature fluctuations on higher-order statistical feature values, environmental drift correction processing is performed on the higher-order statistical feature spatial field of each channel at each time step after spatial registration and before pixel-by-pixel temporal differencing. Specifically, within a reference region where the component is known to be undamaged, the regional mean of the higher-order statistical feature spatial field of each channel at each time step is calculated. The difference between the regional mean of each channel at the current time step and the reference time step is used as the systematic environmental drift estimate for that channel. The systematic environmental drift estimate for the corresponding channel at the corresponding time step is subtracted from the feature values ​​of all pixel positions in the higher-order statistical feature spatial field of each channel at each time step to generate the environmentally corrected multi-time multi-channel higher-order statistical feature spatial field. Subsequently, pixel-by-pixel temporal differencing and gradient calculation are performed on the environmentally corrected multi-time multi-channel higher-order statistical feature spatial field.

[0087] Furthermore, the environmental drift correction process is based on the following principle: systematic factors such as environmental temperature fluctuations have an approximately uniform impact on the magnetic properties of all locations on the component surface, causing an overall additive shift in the spatial distribution of feature values ​​for each channel, without altering the spatial distribution pattern of the feature values. Within a known undamaged reference region, the temporal variation of feature values ​​is entirely caused by environmental factors, and the difference in the regional means of feature values ​​at each time point in the reference region is the estimated value of environmental drift. Subtracting the estimated environmental drift value from the feature values ​​at all pixel locations across the entire field eliminates the global influence of environmental factors while retaining the local feature changes caused by material damage. The reference region should be selected far from known damage areas and expected damage propagation paths, and its area should be large enough to ensure the statistical stability of the regional means.

[0088] Step 5: Perform density clustering on the cross-channel fused gradient vectors at each pixel location to identify the early damage active evolution front and extract the front features;

[0089] The gradient vectors of all channels at each pixel location are concatenated into a fused gradient vector. For the fused gradient vectors at each pixel location on the component surface, cross-channel clustering is performed using the DBSCAN algorithm to identify spatially continuous regions with consistent gradient directions and significant amplitudes. These continuous regions that meet the clustering criteria are marked as early damage active evolution fronts. The principal propagation direction angle, average propagation rate, and front length of each early damage active evolution front are extracted to generate a feature set of early damage spatial evolution fronts for the metal component.

[0090] Wherein, the fused gradient vector G(x,y) at each pixel position (x,y) is the concatenation of gradient vectors from all channels:

[0091] ;

[0092] Where C represents the total number of channels. Let G(x, y) be the two-dimensional gradient vector of the c-th channel at pixel position (x, y), and let the fused gradient vector G(x, y) be a 2C-dimensional vector.

[0093] Furthermore, the fused gradient vector integrates the spatial gradient information of each channel into a unified high-dimensional feature representation. Since different channels correspond to different statistical feature types under different local magnetic field strength windows, the gradient vectors of each channel reflect the spatial evolution characteristics of damage under different physical mechanisms. After concatenating all channel gradient vectors, the fused gradient vector comprehensively characterizes the multi-physical quantity synergistic characteristics of damage evolution at the pixel location in high-dimensional space. The gradient vectors of each channel in the real damage evolution region are consistent in direction, exhibiting a specific directional clustering of the fused gradient vector in high-dimensional space; while spurious changes caused by measurement noise lack directional correlation between channels, resulting in random dispersion of the fused gradient vector in high-dimensional space.

[0094] The DBSCAN algorithm takes the spatial coordinates of each pixel location and the corresponding fused gradient vector as input, and outputs the cluster label for each pixel location. During clustering, the Euclidean distance between the fused gradient vectors and the spatial distance between the pixel spatial coordinates are used as the joint similarity measure. Pixel locations with consistent fused gradient directions and continuous spatial distribution are grouped into the same cluster. Each cluster satisfying the minimum density condition is labeled as an early damage active evolution front. The neighborhood radius and the minimum number of neighborhood points are preset parameters determined based on the resolution of the component surface spatial grid and the minimum scale of the expected damage front.

[0095] Specifically, in the DBSCAN clustering process, for any two pixel positions q1 = (x1, y1) and q2 = (x2, y2) on the surface of a component, the joint similarity metric d(q1, q2) between them is defined as the weighted sum of the spatial distance and the Euclidean distance of the fused gradient vectors:

[0096] d(q1, q2) = α·d spatial (q1, q2) + (1-α)·d gradient (q1, q2);

[0097] in, The spatial Euclidean distance between two pixel locations. Let be the Euclidean distance between the fused gradient vectors of two pixel locations, α∈(0,1) be the weighting coefficient between the spatial distance and the gradient distance, and be a preset parameter. Since the feature values ​​of each channel are dimensionless after Z-score normalization in step 2, the components of the fused gradient vector G(x,y) are also dimensionless. gradient With d in pixels spatialBecause of the different dimensions, the weighting coefficient α is determined based on the ratio between the surface spatial grid resolution of the component and the numerical range of the fused gradient vector, so that the two distances remain comparable in numerical magnitude. When the joint similarity metric distance d(q1, q2) does not exceed the preset neighborhood radius, the two pixel positions are neighbors of each other; when the number of pixels contained in the neighborhood of a pixel position is not less than the preset minimum number of neighborhood points, the pixel position is determined to be a core point; all pixel positions that are reachable from the core point density together form a cluster, that is, an early damage active evolution front.

[0098] Furthermore, the DBSCAN algorithm identifies spatially continuous and feature-similar pixel clusters through density connectivity. For a given pixel location, if the number of neighboring points within its radius reaches the minimum neighboring point threshold, then that location is a core point, indicating that it is located in a feature-dense region. Starting from any core point, all pixel locations within the radius that are feature-similar are gradually connected through density reachability, forming a connected cluster. Since the true damage evolution front is spatially continuous and the fused gradient vector direction is consistent, the joint similarity metric distance between pixel locations within the early active damage evolution front is small, allowing them to aggregate into the same cluster through density connectivity. However, spurious changes caused by measurement noise are spatially isolated or have random fused gradient vector directions, failing to meet the density connectivity condition and being marked as noise points or small-scale outliers and excluded. By adjusting the neighborhood radius and minimum neighboring point parameters, the clustering algorithm's requirements for the continuity and minimum scale of the early active damage evolution front can be controlled.

[0099] The feature extraction method for each early damage active evolution front is as follows: the main expansion direction angle θ is the direction angle after the fusion gradient vector of all pixel positions in the early damage active evolution front is weighted by magnitude; the average expansion rate is the arithmetic mean of the magnitude of the fusion gradient vector of all pixel positions in the early damage active evolution front; the front length is the maximum projection distance of all pixel positions in the early damage active evolution front along the main expansion direction θ.

[0100] Specifically, suppose that a certain early damage active evolution front contains L pixel positions, and the l-th pixel position (x l y l The fusion gradient vector at point (x) is G(x) l y l The magnitude of the fused gradient vector is Where L is the total number of pixel positions included in the early damage active evolution front, and l is the pixel position number within the early damage active evolution front. The principal expansion direction angle θ of the early damage active evolution front is calculated as follows: First, the fused gradient vectors of each pixel position are weighted and summed according to their magnitudes to obtain the weighted composite vector. Then take the weighted composite vector G. weighted The direction angle is taken as the main expansion direction angle θ, that is G weighted , x and G weighted , y The weighted composite vectors G and G are respectively. weighted The components in the x and y directions. The average spreading rate is... The leading edge length is the coordinates of each pixel's position (x, y). l y l The projection value r onto the unit vector (cosθ, sinθ) in the main expansion direction l =x l cosθ+y l The difference between the maximum and minimum values ​​of sinθ, i.e., max l (r) l ) - min l (r) l ).

[0101] Furthermore, the principal expansion direction angle reflects the main expansion direction of the damage front on the component surface. By weighting and summing the fused gradient vectors of each pixel within the early active damage evolution front according to their magnitude, pixels with larger magnitudes contribute more to the principal expansion direction, enabling the principal expansion direction angle to accurately reflect the expansion trend of the most active damage evolution region within the early active damage evolution front. The average expansion rate reflects the overall evolution speed of the early active damage evolution front; a larger average expansion rate indicates that the damage corresponding to the early active damage evolution front develops faster within the current detection period. The front length reflects the spatial span of the damage front in the principal expansion direction; a larger front length indicates that the early active damage evolution front covers a wider spatial range. The principal expansion direction angle, average expansion rate, and front length together describe the spatial morphology and dynamic behavior of the early damage evolution front.

[0102] Step 6: Generate a magnetic coercive damage assessment report for metal components based on the feature set of early damage spatial evolution front of metal components;

[0103] Based on the feature set of early damage spatial evolution fronts of metal components, the spatial coordinates, main propagation direction angle, average propagation rate and front length of each active early damage evolution front on the component surface are summarized and organized to output a magnetic coercive damage assessment report for metal components.

[0104] Furthermore, to predict the evolution trend of early damage and provide graded early warnings, based on obtaining the feature set of early damage spatial evolution fronts of metal components over multiple detection cycles, the following steps are also included: Linear regression trend extrapolation is performed on the main expansion direction angle and average expansion rate of each active early damage evolution front in multiple detection cycles to predict the expected spatial location and expected coverage of each active early damage evolution front in the next detection cycle. The input to the linear regression is the position coordinate sequence and expansion rate time sequence of each active early damage evolution front in historical detection cycles, and the output is the predicted position coordinates and predicted expansion rate for the next detection cycle. The predicted location of each active early damage evolution front is superimposed and compared with the spatial coordinates of the critical load-bearing area of ​​the component to calculate the expected time for each active early damage evolution front to reach the nearest critical load-bearing area boundary at the current expansion rate. Active early damage evolution fronts with expected arrival times lower than a preset safe time threshold are marked as high-risk early warning objects, and corresponding graded early warning information is output in the magnetic coercive damage assessment report of the metal component. The graded early warning information includes the risk level of each active early damage evolution front, the expected time to reach the critical load-bearing area, and the corresponding critical load-bearing area identifier.

[0105] Furthermore, linear regression trend extrapolation establishes a time series prediction model based on the evolution trajectory of each early damage active evolution front across multiple historical detection cycles. For a specific early damage active evolution front, the center position coordinates and average expansion rate of each early damage active evolution front at each historical detection time are extracted. Linear regression models are established with the detection time as the independent variable and either the center position coordinates or the average expansion rate as the dependent variable. The parameters of the linear regression models are determined by fitting historical data using the least squares method. The expected time of the next detection cycle is substituted into the linear regression model to obtain the predicted center position coordinates and predicted average expansion rate of the early damage active evolution front in the next cycle. Combining the predicted main expansion direction angle and front length, the expected coverage area of ​​the early damage active evolution front in the next cycle is calculated. The expected coverage area is geometrically superimposed with the spatial coordinates of the critical load-bearing area of ​​the component to calculate the spatial distance from the boundary of the early damage active evolution front to the boundary of the nearest critical load-bearing area. Dividing this distance by the average expansion rate yields the expected arrival time.

[0106] According to an embodiment of this method, higher-order statistical morphological characteristics (local skewness, local kurtosis, and inter-cycle microvariation) of the local region of the hysteresis loop are used instead of the absolute value of coercivity as the basic physical quantity for damage assessment. When early microstructural changes occur in metallic materials, the formation of dislocation substructures alters the pinning characteristics of local magnetic domain walls, leading to minute distortions in the morphology of the hysteresis loop within a specific magnetic field strength range. Although these morphological distortions are insufficient to cause significant changes in the absolute value of coercivity, they do cause changes in the skewness and kurtosis of the magnetization distribution within the local region, while also increasing the fluctuations in the magnetization response between repeated magnetization cycles. Therefore, higher-order statistics can provide an effective numerical response to the minute morphological distortions of the hysteresis loop that are masked by measurement noise in the absolute value of coercivity, thus overcoming the problem of insufficient sensitivity of the absolute value of coercivity method in the early microscopic damage stage.

[0107] Furthermore, this method extends the aforementioned higher-order statistical features from single-point independent detection to multi-channel spatial fields. By performing temporal difference analysis on the multi-channel higher-order statistical feature spatial fields at multiple time points, the method obtains the higher-order statistical feature change fields of each channel and calculates the spatial gradient vector of the higher-order statistical feature change fields. Then, density clustering of cross-channel fused gradient vectors is used to identify regions with consistent gradient directions and spatial continuity. Since the true early damage evolution is spatially manifested as a continuous region that extends along a specific direction, the multi-channel gradient vectors of the true early damage evolution are consistent in direction. However, spurious changes caused by measurement noise are randomly distributed in space and lack directional consistency between channels. Therefore, cross-channel gradient vector field clustering can separate the true damage evolution front from the noise background, overcoming the problem of indistinguishability between true signals and spurious changes in multi-channel higher-order statistical feature spatial field difference analysis. This simultaneously meets the evaluation requirements of early high-sensitivity damage detection and damage spatial evolution tracking.

[0108] The high-pressure reactor shell (made of low-alloy high-strength steel 15CrMoR) in a petrochemical plant is at risk of early fatigue damage near the heat-affected zone of the weld seam under long-term alternating loads and corrosive media coupling. This reactor has been in continuous service for several years and has entered its periodic non-destructive testing cycle. The testing system divides the shell surface into a spatial grid at 10mm intervals, setting a total of 180 measurement points (15×12). Four fixed marker points are set on the shell surface for spatial registration between multiple tests. The current test is the third periodic test (denoted as time t=3), and historical test data from t=1 and t=2 are available. The number of repeated magnetization cycles is set to N=25, the number of local interval windows is set to w=5, and each test generates a high-order statistical characteristic spatial field with 3w=15 channels.

[0109] At the detection time t=3, the magnetization acquisition device performed 25 repeated standardized magnetization cycles on each of the 180 measurement points. Taking measurement point p47 near the heat-affected zone of the weld as an example, the spatial coordinates of this point are (x=47mm, y=83mm), and the magnetic field strength sampling points cover -8000A / m to +8000A / m, with a total of 401 sampling points (j=1, 2, ..., 401). Taking the magnetic field strength sampling point H201=0A / m near the coercivity of the rising support as an example, the original magnetization strength acquired after 25 cycles is shown below.

[0110] Table 1. Original values ​​of magnetization intensity at measurement point p47 and sampling point H201 after 25 cycles:

[0111] ;

[0112] Perform synchronous averaging on the above 25 original magnetization values:

[0113] .

[0114] The standard deviation of random noise in a single measurement is approximately 2.8 A / m. After averaging 25 synchronous measurements, the standard deviation of noise decreases to... The signal-to-noise ratio was significantly improved. The above synchronous averaging process was repeated for all 401 sampling points of 180 measurement points to generate high signal-to-noise ratio average hysteresis loop data for each measurement point.

[0115] The magnetic field strength of the ascending branch spans from 0 A / m to 8000 A / m, divided into 5 equal-width local interval windows, each 1600 A / m wide, containing 40 magnetic field strength sampling points (Q). Taking measurement point p47 as an example, three types of higher-order statistical features are extracted from its third window (magnetic field strength range of 3200 A / m to 4800 A / m, corresponding to the sensitive section near the coercivity).

[0116] The average magnetization value corresponding to the 40 sampling points within this window was calculated as follows: mean Standard deviation .

[0117] Local skewness value: ;

[0118] Local kurtosis values: .

[0119] Intercycle variance: The mean magnetization intensity of each of the 25 original hysteresis loops is calculated within the third window, and then the variance across cycles is calculated. .

[0120] After performing the above three types of calculations on 180 measurement points in 5 windows, Z-score standardization was performed on the three types of feature values ​​of each channel for all measurement points to eliminate dimensional differences. The original value before standardization and the standardized value after standardization of point p47 in all 5 windows are shown as an example below.

[0121] Table 2. Original values ​​and Z-score standardized values ​​of higher-order statistical features for each window at measurement point p47:

[0122] .

[0123] It can be seen that the standardized values ​​of the three types of features at point p47 in the third window (the section near the coercivity) are all significantly higher, indicating that the measurement point has the characteristics of skewed statistical distribution of domain flipping, sharpened distribution and enhanced intercycle fluctuation in this magnetic field strength section, suggesting that early microscopic damage may be concentrated in the dislocation activity mechanism corresponding to this section.

[0124] The 15 channel feature values ​​of the 180 measurement points, after Z-score normalization, are arranged into discrete point sets according to the two-dimensional coordinates (x, y) of each point on the unfolded surface of the cylinder. Taking the skewness channel of the third window (channel number 3, corresponding to the normalized field of window 3 skewness) as an example, the spatial coordinates and normalized feature values ​​of 9 measurement points near the heat-affected zone of the weld are selected as input control points for Kriging interpolation.

[0125] Table 3. Kriging interpolation input control point data for the third window skewness channel (at time t=3):

[0126] .

[0127] Kriging interpolation uses the aforementioned 9 control points (and the remaining 171 measurement points) as input. A variogram model is established using the spatial autocorrelation function, and the Kriging equations are solved to obtain the interpolation weights. Estimation is performed on all pixel positions within a uniform 1mm resolution grid (approximately 150×120 pixels) on the cylinder's unfolded surface to generate a continuous spatial field for the skewness channel of the third window. After completing the Kriging interpolation for each of the 15 channels, a complete multi-channel high-order statistical characteristic spatial field at time t=3 is obtained.

[0128] Before performing temporal differencing, the spatial fields of the multi-channel high-order statistical features at times t=1, t=2, and t=3 are first spatially registered. Using four fixed marker points on the cylinder surface (coordinates (10, 10), (140, 10), (140, 110), (10, 110) mm) as control points, an affine transformation is performed on the spatial fields at each time point to align them to a unified reference coordinate system. After registration, the mean values ​​of each channel region are calculated for a known undamaged reference region (the lower right corner region away from the weld, x∈[120, 140] mm, y∈[95, 110] mm). After estimating the environmental drift and subtracting it from the entire field, pixel-by-pixel temporal differencing is performed at times t=3 and t=2.

[0129] Taking the skewness channel of the third window at pixel position (45, 83) (corresponding to measurement point p47) and its spatial neighboring pixels as an example, the local values ​​of the differential field are listed.

[0130] Table 4. Local values ​​of time difference in the third window skew channel (after environmental correction):

[0131] .

[0132] For pixel position (45, 83), calculate the spatial gradient vector using the center difference:

[0133] ;

[0134] .

[0135] Since p47 is located at the extreme center of the difference field, the partial derivatives in both directions at this point are close to zero, while there is a gradient dominated by the weld direction (along the y-axis) in its surrounding area (e.g., from (43, 83) to (47, 83)). After calculating the gradient vector field for each of the 15 channels, the fusion clustering step is performed.

[0136] The 15 channel gradient vectors at each pixel location are concatenated into a 30-dimensional fused gradient vector. For the 1mm resolution grid of the cylinder unfolded surface in this case, the DBSCAN neighborhood radius is set to 8 (a hybrid metric combining spatial and gradient distance), the minimum number of neighborhood points is 6, and the spatial-gradient weight coefficient α = 0.55.

[0137] The clustering results identified one significant cluster in the heat-affected zone of the weld, labeled as the early damage active evolution front F-01, which contains L=37 pixel locations, concentrated in the region (x∈[39,57]mm, y∈[77,91]mm) and extends along the weld direction (approximately the y-axis direction).

[0138] The magnitude-weighted composite vector of the fused gradient vectors at the 37 pixel positions within the leading edge F-01 is calculated, yielding an x-axis component of 3.12 and a y-axis component of 18.74. Therefore, the main expansion direction angle is:

[0139] .

[0140] This indicates that the damage front extends along a direction close to the positive y-axis (i.e., parallel to the weld direction), which is consistent with the typical propagation pattern of fatigue crack initiation in the heat-affected zone of this type of pressure vessel.

[0141] Average expansion rate: (Dimensionless normalized feature change rate / detection cycle).

[0142] Leading edge length: The projection of each pixel's position coordinates onto the unit vector (cos80.5°, sin80.5°) = (0.165, 0.986) in the main expansion direction. The maximum projection value is approximately 0.165×47+0.986×91=7.76+89.73=97.49mm, and the minimum projection value is approximately 0.165×39+0.986×77=6.44+75.92=82.36mm. The leading edge length is 97.49-82.36=15.1mm.

[0143] By combining the F-01 characteristic data of the leading edge at three detection times t=1, t=2, and t=3, linear regression trend extrapolation is performed on the main expansion direction angle and average expansion rate to predict the leading edge state at t=4, and spatial superposition analysis is performed with the key bearing area of ​​the cylinder (the main stress concentration zone is located at y=105mm).

[0144] Table 5 Historical data and predicted values ​​of the multi-period evolution characteristics of the F-01 front:

[0145] .

[0146] Using the detection time t as the independent variable, a linear regression was established on the y-coordinate of the leading edge center: the slope was approximately 3.7 mm / cycle, the intercept was approximately 73.7 mm, and the fit was good. The predicted y-coordinate of the leading edge center at t=4 is approximately 88.2 mm, about 16.8 mm away from the boundary of the critical bearing area (y=105 mm). Based on the current average expansion rate and considering the expansion component of the leading edge along the y-direction, the expected arrival time is approximately 4.4 detection cycles. This value is lower than the preset safe time threshold (6 detection cycles), and leading edge F-01 is marked as a high-risk warning target. The following graded warning information is output in the assessment report:

[0147] Risk level: High risk;

[0148] Front marking: F-01 (Y-axis extension of the heat-affected zone of the weld);

[0149] Expected time to reach the critical load-bearing area: approximately 4.4 detection cycles;

[0150] Corresponding key load-bearing area marker: Principal stress concentration zone KZ-01 (at y=105mm);

[0151] The data flow logic of the entire implementation process is as follows: Step 1: By repeatedly acquiring and averaging data 25 times, the standard deviation of the noise of the original hysteresis loop is reduced to 1 / 5 of its original value, laying a high-quality data foundation for subsequent extraction of higher-order statistical features; Step 2: Skewness, kurtosis, and inter-cycle microvariation are extracted from the five local windows of the average hysteresis loop, and the dimensions are unified by Z-score standardization. The significantly high value of the third window at point p47 first reveals the early damage signal at this measurement point; Step 3: Kriging interpolation expands the 180 discrete measurement points into a continuous 15-channel spatial field covering the unfolded surface of the cylinder, giving the damage signal spatial resolution; Step 4: The data flow logic of the hysteresis loop at t=3 and t= Step 2 involves spatial registration, environmental drift correction, and pixel-by-pixel differencing, followed by the calculation of the gradient vector field to transform the temporal evolution of damage into directional spatial gradient information. Step 5 uses DBSCAN cross-channel clustering to identify the leading edge F-01 from the 30-dimensional fused gradient vector, extracting quantitative leading edge features such as the main expansion direction angle of 80.5°, the average expansion rate of 0.671, and the leading edge length of 15.1 mm. Step 6 uses linear regression extrapolation based on three historical cycles to predict the position of the leading edge in the fourth cycle, calculating the expected arrival time of only 4.4 cycles from the critical bearing area, triggering a high-risk warning, and completing the full-link data flow from raw magnetization data to quantitative damage evolution assessment and risk warning.

[0152] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for assessing magnetic coercive damage to metal components, characterized in that, Includes the following steps: Multiple standardized magnetization cycles are performed on each measurement point on the surface of the metal component according to a preset spatial grid to acquire multiple hysteresis loop raw data of each measurement point. The multiple hysteresis loop raw data of each measurement point are synchronously averaged to generate high signal-to-noise ratio average hysteresis loop data of each measurement point. For the rising branch data of the high signal-to-noise ratio average hysteresis loop at each measurement point, multiple local interval windows are uniformly divided on the magnetic field strength axis. The local skewness value and local kurtosis value are calculated for the magnetization intensity data in each local interval window. Based on the original data of multiple hysteresis loops at each measurement point, the inter-cycle microvariability value is calculated in each local interval window, generating the local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariability value sequence for each measurement point. The local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariation value sequence of each measurement point are arranged according to the two-dimensional spatial coordinates of each measurement point on the surface of the component. Kriging interpolation is performed on each channel to generate a multi-channel high-order statistical feature space field at the current detection time. The system obtains the multi-channel high-order statistical feature spatial field of historical detection time, calculates the pixel-by-pixel difference of the high-order statistical feature spatial field of each channel between the current detection time and the adjacent historical detection time to obtain the high-order statistical feature change field of each channel, calculates the spatial gradient vector of each pixel position in the high-order statistical feature change field of each channel, and generates the high-order statistical feature change gradient vector field of each channel. All channel gradient vectors at each pixel location are concatenated into a fused gradient vector. Density clustering is performed on the fused gradient vectors at each pixel location on the component surface to identify spatially continuous regions with consistent gradient directions and significant amplitudes, which are then marked as active evolution fronts of early damage. Front features of each active evolution front of early damage are extracted to generate a feature set of spatial evolution fronts of early damage in metal components. A magnetic coercive damage assessment report for metal components is generated based on the feature set of early damage spatial evolution front of the metal components.

2. The method for assessing magnetic coercive damage to metal components according to claim 1, characterized in that, The synchronous averaging process refers to taking the arithmetic mean of the magnetization values ​​of multiple hysteresis loops at the same measurement point at the same magnetic field strength sampling point; the standardized magnetization cycle refers to applying a periodically changing magnetic field with a preset waveform to the measurement point, so that the material undergoes a complete magnetization process from negative saturation magnetization state to positive saturation magnetization state and back to negative saturation magnetization state. The standardized magnetization cycle of each measurement point adopts the same magnetic field strength variation range, the same magnetic field variation rate and the same sampling time interval.

3. The method for assessing magnetic coercive damage to metal components according to claim 1, characterized in that, The local skewness value is calculated as follows: for the average magnetization value corresponding to each magnetic field strength sampling point within the local interval window, calculate the ratio of the cube of the magnetization deviation to the cube of the standard deviation of the magnetization. The local kurtosis value is calculated as follows: for the average magnetization value corresponding to each magnetic field strength sampling point within the local interval window, the ratio of the fourth power of the magnetization deviation to the fourth power of the magnetization standard deviation is calculated; the inter-cycle microvariability value is calculated as follows: for multiple hysteresis loops at the same measurement point, the mean magnetization value of each hysteresis loop within the local interval window is calculated, and then the variance between the multiple mean magnetization values ​​is calculated.

4. The method for assessing magnetic coercive damage to metal components according to claim 1, characterized in that, Before performing Kriging interpolation on each channel, Z-score standardization is performed on the local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariance value sequence of each measurement point to unify the feature values ​​of each channel into dimensionless values ​​with a mean of zero and a standard deviation of one. Each channel in the multi-channel high-order statistical feature space field corresponds to a statistical feature type under a local interval window number, and the total number of channels is three times the number of local interval windows.

5. The method for assessing magnetic coercive damage to metal components according to claim 1, characterized in that, Before calculating the pixel-by-pixel difference, spatial registration processing is performed on the multi-channel high-order statistical feature space field at all detection times. The spatial registration processing includes: using the fixed marker points preset on the surface of the component as control points, selecting at least three non-collinear fixed marker points, identifying the actual pixel coordinates of the control points in the multi-channel high-order statistical feature space field at each time, establishing a correspondence with the reference coordinates of the corresponding control points in a unified reference coordinate system, determining the affine transformation matrix by solving a system of linear equations, applying the affine transformation matrix to all pixel positions in the multi-channel high-order statistical feature space field at each time for geometric alignment, and obtaining the feature values ​​of the mapped pixel positions through bilinear interpolation to generate the registered multi-channel high-order statistical feature space field.

6. The method for assessing magnetic coercive damage of metal components according to claim 5, characterized in that, After spatial registration and before pixel-by-pixel differencing, environmental drift correction is performed on the high-order statistical feature spatial field of each channel at each time. The environmental drift correction includes: calculating the regional mean of the high-order statistical feature spatial field of each channel at each time within a reference area where the component is known to be undamaged; using the difference between the regional mean of each channel at the current time and the reference time as the systematic environmental drift estimate of that channel; and subtracting the systematic environmental drift estimate of the corresponding channel at the corresponding time from the feature value of all pixel positions in the high-order statistical feature spatial field of each channel at each time to generate the environmentally corrected multi-time multi-channel high-order statistical feature spatial field.

7. The method for assessing magnetic coercive damage to metal components according to claim 1, characterized in that, The density clustering uses the DBSCAN algorithm. During the clustering process, the joint similarity metric is the weighted sum of the Euclidean distance between the fused gradient vectors and the spatial distance between the pixel spatial coordinates. The joint similarity metric is the sum of the spatial Euclidean distance and the fused gradient vector Euclidean distance multiplied by their respective weight coefficients, where the sum of the spatial distance weight coefficient and the gradient distance weight coefficient is one. When the joint similarity metric between two pixel positions does not exceed the preset neighborhood radius, the two pixel positions are neighbors of each other. When the number of pixels contained in the neighborhood of a pixel position is not less than the preset minimum number of neighborhood points, the pixel position is determined to be a core point. All pixel positions that are density-reachable from the core point together form a cluster, and each cluster that meets the minimum density condition is marked as an early damage active evolution front.

8. The method for assessing magnetic coercive damage of metal components according to claim 1, characterized in that, The leading edge features include the main expansion direction angle, the average expansion rate, and the leading edge length; the main expansion direction angle is the direction angle of the weighted composite vector obtained by summing the fused gradient vectors of all pixel positions within the early damage active evolution leading edge according to their magnitude. The average expansion rate is the arithmetic mean of the magnitudes of the fused gradient vectors at all pixel locations within the active evolution front of early damage; the front length is the difference between the maximum and minimum values ​​of the projections of the coordinates of all pixel locations within the active evolution front of early damage onto the unit vector in the main expansion direction.

9. The method for assessing magnetic coercive damage of metal components according to claim 1, characterized in that, Based on acquiring the feature set of early damage spatial evolution fronts of metal components across multiple detection cycles, the method further includes: performing linear regression trend extrapolation on the main expansion direction angle and average expansion rate of each early damage active evolution front in multiple detection cycles to predict the expected spatial location and expected coverage of each early damage active evolution front in the next detection cycle; superimposing and comparing the predicted location of each early damage active evolution front with the spatial coordinates of the critical load-bearing area of ​​the component to calculate the expected time for each early damage active evolution front to reach the nearest critical load-bearing area boundary at the current expansion rate; marking early damage active evolution fronts with expected arrival times lower than a preset safe time threshold as high-risk warning objects, and outputting corresponding graded warning information in the metal component magnetic coercive damage assessment report, wherein the graded warning information includes the risk level of each early damage active evolution front, the expected time to reach the critical load-bearing area, and the corresponding critical load-bearing area identifier.

10. A magnetic coercive damage assessment system for metal components, used to perform the magnetic coercive damage assessment method for metal components according to any one of claims 1 to 9, characterized in that, include: The hysteresis loop acquisition and averaging module is used to perform multiple repeated standardized magnetization cycles to acquire raw data of multiple hysteresis loops at each measurement point on the surface of the metal component, and to perform synchronous averaging of the raw data of multiple hysteresis loops at each measurement point to generate high signal-to-noise ratio averaged hysteresis loop data for each measurement point. The advanced statistical feature extraction module is used to uniformly divide the rising branch data of the high signal-to-noise ratio average hysteresis loop at each measurement point into multiple local interval windows on the magnetic field strength axis. It calculates the local skewness value and local kurtosis value for the magnetization intensity data in each local interval window, and calculates the inter-cycle microvariability value in each local interval window based on the original data of multiple hysteresis loops at each measurement point, generating the local skewness value sequence, local kurtosis value sequence, and inter-cycle microvariability value sequence for each measurement point. The spatial field generation module is used to arrange the local skewness value sequence, local kurtosis value sequence and inter-cycle microvariation value sequence of each measurement point according to the two-dimensional spatial coordinates of each measurement point on the surface of the component, and perform Kriging interpolation on each channel to generate a multi-channel high-order statistical feature spatial field at the current detection time. The temporal difference and gradient calculation module is used to obtain the multi-channel high-order statistical feature spatial field of historical detection time. It calculates pixel-by-pixel difference of the high-order statistical feature spatial field of each channel between the current detection time and the adjacent historical detection time to obtain the high-order statistical feature change field of each channel. It also calculates the spatial gradient vector of each pixel position in the high-order statistical feature change field of each channel to generate the high-order statistical feature change gradient vector field of each channel. The density clustering and front identification module is used to concatenate all channel gradient vectors at each pixel position into a fused gradient vector, perform density clustering on the fused gradient vectors at each pixel position on the component surface, identify spatially continuous regions with consistent gradient directions and significant amplitudes and mark them as active evolution fronts of early damage, extract the front features of each active evolution front of early damage, and generate a feature set of spatial evolution fronts of early damage in metal components. The assessment report generation module is used to generate a magnetic coercive damage assessment report for the metal component based on the feature set of the early damage spatial evolution front of the metal component.

Images