Intelligent Diagnosis Method and System for Material Damage Based on Magnetic Characteristic Data Fusion

Abstract

The present invention provides a method and system for intelligent diagnosis of material damage by magnetic characteristic data fusion, which relates to the field of nondestructive testing technology. The method includes using a magnetic stress detection robot to collect magnetic characteristic data; performing multi-scale decomposition on the data to construct a magnetic domain evolution feature spectrum; inputting the feature spectrum into a composite neuron network to extract a material damage feature matrix; constructing a fusion feature vector by using tensor dimensionality reduction and a multi-layer belief transfer network; and realizing damage calculation through temperature field modulation and adaptive weight allocation to obtain the damage level and remaining life. The present invention realizes the accurate diagnosis of material damage in the weld area of a pressure vessel.

Description

Technical Field

[0001] The present invention relates to the technical field of nondestructive testing, and particularly to an intelligent diagnosis method and system for material damage by fusing magnetic characteristic data. Background Art

[0002] As a key device in industrial production, the safe operation of a pressure vessel is directly related to production safety and economic benefits. During long-term service, the weld area of a pressure vessel is prone to fatigue, creep and other damages due to complex stress and temperature environments, resulting in the degradation of structural performance and even failure. Therefore, it is of great significance to accurately diagnose material damage and predict the life of the weld area of a pressure vessel. Magnetic characteristic characterization is widely regarded as an effective means to evaluate the changes in the microstructure and damage state of materials. Among them, characteristic parameters such as magnetic coercivity, remanent magnetization intensity and magnetic permeability are closely related to the microstructure and damage degree of materials. In recent years, researchers have developed a variety of material damage diagnosis methods based on magnetic characteristic data. By collecting magnetic characteristic data of materials and analyzing the relationship between them and the material damage state, the prediction of the remaining life of materials is realized.

[0003] However, the existing magnetic characteristic-based material damage diagnosis methods have defects and deficiencies. The existing methods mainly focus on the correlation research between a single magnetic characteristic parameter and material damage, lacking the fusion analysis of multiple magnetic characteristic data, resulting in insufficient comprehensiveness and accuracy of the diagnosis results, especially in the evaluation of fatigue-creep coupling damage under complex stress environments; traditional diagnosis methods mostly stay at the macroscopic statistical level in the processing of magnetic characteristic data, failing to deeply explore the microscopic mechanism information of magnetic domain wall migration and magnetic domain evolution contained in the data, and it is difficult to accurately reflect the internal relationship between the evolution process of the material microstructure and the damage degree; the existing intelligent diagnosis models generally adopt a single neural network structure, lacking the ability of multi-scale decomposition and hierarchical processing of magnetic characteristic features, and unable to effectively handle the influence of temperature field changes on material damage, thus limiting the accuracy of damage diagnosis and life prediction, especially in the reliability evaluation under long-term high-temperature service conditions, there are obvious deficiencies. Summary of the Invention

[0004] The embodiments of the present invention provide an intelligent diagnosis method and system for material damage by fusing magnetic characteristic data, which can solve the problems in the prior art.

[0005] In the first aspect of the embodiments of the present invention, an intelligent diagnosis method for material damage by fusing magnetic characteristic data is provided, including:

[0006] Based on the weld area of a pressure vessel, use a magnetic stress detection robot equipped with a magnetic coercivity detection probe to collect magnetic characteristic data, including magnetic coercivity data, remanent magnetization intensity data and magnetic permeability data;

[0007] Perform multi-scale decomposition on the magnetic property data, construct a magnetic domain evolution characteristic spectrum through the band coefficients with the maximum energy entropy, and obtain a magnetic domain wall migration characteristic map;

[0008] Based on the magnetic domain wall migration characteristic map, input the magnetic domain evolution characteristic spectrum into a composite neural network, extract the magnetic domain structure characteristics, grain orientation characteristics, and texture evolution characteristics of the material, and obtain a material damage characteristic matrix;

[0009] In the material damage characteristic matrix, use tensor dimensionality reduction transformation and the hierarchical training mechanism of a multi-layer belief transfer network to construct a magnetic property fusion feature vector;

[0010] Based on the magnetic property fusion feature vector, perform fatigue-creep coupling damage calculation through temperature field modulation and adaptive weight allocation, and combine probability density analysis and an adaptive kernel function to obtain the material damage level and remaining life value.

[0011] In an alternative embodiment, performing multi-scale decomposition on the magnetic property data, constructing a magnetic domain evolution characteristic spectrum through the band coefficients with the maximum energy entropy, and obtaining a magnetic domain wall migration characteristic map includes:

[0012] Obtain a band coefficient matrix by performing multi-layer wavelet packet decomposition on the collected magnetic property data, and the band coefficient matrix characterizes the multi-scale characteristic information during the magnetization process;

[0013] Calculate the energy entropy value of the band coefficient matrix to construct an energy distribution characteristic spectrum, identify the main band magnetization response in the internal stress concentration region of the material based on the energy distribution characteristic spectrum, and select the main band magnetization response as the characteristic band;

[0014] Reconstruct the magnetization curve according to the time-frequency evolution law of the characteristic band, perform magnetic domain wall pinning-unpinning motion analysis on the magnetization curve, and obtain the magnetic domain flipping kinetic density function through spectral analysis;

[0015] Construct a magnetic domain order parameter distribution matrix based on the main band component of the magnetic domain flipping kinetic density function, which contains the anisotropic characteristics of damage evolution;

[0016] Perform coupled calculation on the magnetic domain order parameter distribution matrix and the internal dislocation density distribution of the material to obtain a two-dimensional characteristic function, and generate a magnetic domain wall migration characteristic map based on the two-dimensional characteristic function.

[0017] In an alternative embodiment, performing magnetic domain wall pinning-unpinning motion analysis on the magnetization curve and obtaining the magnetic domain flipping kinetic density function through spectral analysis includes:

[0018] Calculate the local magnetic susceptibility of the reconstructed magnetization curve, and divide the magnetization curve into a pinning-dominated region and an unpinning-dominated region based on the comparison result between the local magnetic susceptibility and a preset critical magnetic susceptibility;

[0019] Establish a multi-pinning point potential field distribution function in the pinning-dominated region, including the pinning potential field strength coefficient, the pinning point position coordinates, and the pinning potential field action range parameter; construct a magnetic domain wall migration velocity field function in the depinning-dominated region, including the characteristic velocity parameter, the local effective magnetic field, and the critical magnetic field strength;

[0020] Substitute the multi-pinning point potential field distribution function and the magnetic domain wall migration velocity field function into the pinning-depinning coupling equation, and solve to obtain the magnetic domain wall displacement response function, including the amplitude attenuation coefficient, the natural oscillation frequency, and the initial phase;

[0021] Perform autocorrelation operation on the magnetic domain wall displacement response function and perform Fourier transform to obtain the kinetic density function, which characterizes the transition characteristics of the magnetic domain wall between pinning points.

[0022] In an alternative embodiment, based on the magnetic domain wall migration feature map, input the magnetic domain evolution feature spectrum into a composite neural network, extract the magnetic domain structure features, grain orientation features, and texture evolution features of the material, and obtain a material damage feature matrix including:

[0023] Input the magnetic domain evolution feature spectrum into a composite neural network including a parallel convolutional layer, where the parallel convolutional layer is provided with convolutional kernels of different scales, separate the feature components of the magnetic domain wall migration through the parallel convolutional layer, and obtain a magnetic domain structure feature vector through the feature enhancement layer;

[0024] Perform a non-linear transformation on the magnetic domain evolution feature spectrum in the feature mapping layer of the composite neural network, extract the spatial distribution information of the grain arrangement, and generate a grain orientation feature vector;

[0025] Separate the texture information in the magnetic domain evolution feature spectrum through the feature decoupling layer of the composite neural network, and obtain a texture evolution feature vector using a dynamic feature extraction unit;

[0026] Assign weight coefficients to the magnetic domain structure feature vector, the grain orientation feature vector, and the texture evolution feature vector according to the information entropy criterion, establish a feature interaction matrix to calculate the correlation degree, and obtain a multi-feature fusion vector;

[0027] Input the multi-feature fusion vector into a damage feature mapping network, perform feature transformation through a residual connection structure and a non-linear activation function, and output a material damage feature matrix.

[0028] In an alternative embodiment, in the material damage feature matrix, use tensor dimensionality reduction transformation and the hierarchical training mechanism of a multi-layer belief transfer network to construct a magnetic property fusion feature vector including:

[0029] Reconstruct the material damage feature matrix into a third-order tensor through tensor outer product operation, and the third-order tensor is determined by tensor decomposition factors and tensor rank to form a damage feature tensor;

[0030] Perform Tucker decomposition on the damage feature tensor, construct the tensor product of the core tensor and each modal projection matrix, and obtain a dimensionality-reduced feature tensor;

[0031] Establish node state vectors in a multi-layer belief propagation network, calculate belief propagation values between nodes based on connection weights, and form a network belief state matrix;

[0032] Input the dimensionality-reduced feature tensor into the multi-layer belief propagation network, construct a hierarchical loss function based on a reconstruction loss function, a belief propagation loss function, and a consistency constraint loss function, and perform hierarchical training;

[0033] According to the gradient information of the hierarchical loss function, use a learning rate parameter and a regularization parameter to iteratively update the network belief state matrix, generate a feature weight matrix, and extract global features and local features based on the feature weight matrix;

[0034] Based on the feature weight matrix, perform weighted combination on the global feature and the local feature, and obtain a magnetic property fusion feature vector through an activation function.

[0035] In an alternative embodiment, based on the magnetic property fusion feature vector, fatigue-creep coupling damage calculation is realized through temperature field modulation and adaptive weight allocation, and combined with probability density analysis and an adaptive kernel function, the material damage level and remaining life value are obtained, including:

[0036] Extract the magnetic parameter group in the magnetic property fusion feature vector, perform non-linear mapping on the magnetic parameter group through a temperature field modulation function, establish the correlation law between the magnetic field and the stress field, and calculate the fatigue damage variable;

[0037] Pass the magnetic property fusion feature vector through an adaptive weight allocation algorithm to determine the creep stress correction coefficient, and calculate the creep damage variable in combination with the creep strain evolution law;

[0038] Perform coupling processing on the fatigue damage variable and the creep damage variable, calculate the field effect superposition amount using a non-linear coupling function, establish a damage accumulation path, and obtain the total damage amount;

[0039] Use piecewise probability density calculation to analyze the distribution characteristics of the total damage amount, and use an adaptive kernel function to calculate the two-parameter life distribution probability value to realize the dynamic assessment of the material damage state;

[0040] Set the hierarchical damage assessment criteria according to the two-parameter lifetime distribution probability value, determine the multi-level damage discrimination boundary in combination with the material damage critical threshold, and calculate the material damage level.

[0041] Substitute the total damage amount and the material damage level into the damage evolution rate equation, and determine the law of life attenuation through integral calculation, and output the remaining life value of the material.

[0042] In an optional embodiment, perform coupling processing on the fatigue damage variable and the creep damage variable, calculate the field effect superposition amount using a non-linear coupling function, establish a damage accumulation path, and obtain the total damage amount including:

[0043] Establish a fatigue-creep damage state space, map the fatigue damage variable and the creep damage variable in the state space, and determine the damage state point.

[0044] Construct a damage interaction gain function, and calculate the promotion effect of fatigue damage on creep damage and the acceleration effect of creep damage on fatigue damage based on the damage state point.

[0045] Use the damage interaction gain function to non-linearly couple the fatigue damage variable and the creep damage variable, and calculate the field effect superposition amount.

[0046] Track the evolution trajectory of the damage state point in the state space, establish a damage accumulation path, and the damage accumulation path reflects the development law of material damage.

[0047] Calculate the damage evolution rate based on the damage accumulation path, and obtain the total damage amount in combination with the critical damage threshold.

[0048] In the second aspect of the embodiments of the present invention, a material damage intelligent diagnosis system for magnetic characteristic data fusion is provided, including:

[0049] The first unit is used to collect magnetic characteristic data based on the weld area of the pressure vessel, using a magnetic stress detection robot equipped with a magnetic coercivity detection probe, including magnetic coercivity data, remanent magnetization intensity data, and magnetic permeability data.

[0050] The second unit is used to perform multi-scale decomposition on the magnetic characteristic data, construct a magnetic domain evolution characteristic spectrum through the frequency band coefficient with the maximum energy entropy, and obtain a magnetic domain wall migration characteristic map.

[0051] The third unit is used to input the magnetic domain evolution characteristic spectrum into a composite neuron network based on the magnetic domain wall migration characteristic map, extract the magnetic domain structure characteristics, grain orientation characteristics, and texture evolution characteristics of the material, and obtain a material damage characteristic matrix.

[0052] The fourth unit is configured to construct a magnetic property fusion feature vector in the material damage feature matrix by using tensor dimensionality reduction transformation and the hierarchical training mechanism of a multi-layer belief transfer network;

[0053] The fifth unit is configured to perform fatigue-creep coupling damage calculation based on the magnetic property fusion feature vector through temperature field modulation and adaptive weight allocation, and combine probability density analysis and an adaptive kernel function to obtain the material damage level and the remaining life value.

[0054] In a third aspect of the embodiments of the present invention, there is provided an electronic device, including:

[0055] A processor;

[0056] A memory for storing instructions executable by the processor;

[0057] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0058] In a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0059] In the embodiments of the present invention, by collecting multi-dimensional magnetic property data through a magnetic stress detection robot and constructing a magnetic domain evolution feature spectrum by using a multi-scale decomposition technology, the internal microstructural changes of the material can be comprehensively reflected, and the sensitivity and accuracy of material damage detection are improved; by extracting multi-level material features through a composite neuron network, combining tensor dimensionality reduction transformation and a multi-layer belief transfer network to construct a fusion feature vector, the efficient fusion of magnetic property data is realized, the robustness of damage recognition is enhanced, and the limitation of single-parameter detection is effectively avoided; by using temperature field modulation and adaptive weight allocation technology to perform fatigue-creep coupling damage calculation, combining probability density analysis and an adaptive kernel function to evaluate the damage level and the remaining life, a scientific and reliable prediction basis is provided for the safe operation of pressure vessels, and it has strong practical value and promotion significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a schematic flowchart of the intelligent diagnosis method for material damage by magnetic property data fusion according to the embodiments of the present invention;

[0061] Figure 2 It is a graph of experimental results of local magnetic susceptibility analysis and magnetization curve region division;

[0062] Figure 3 It is a three-dimensional visualization graph of the damage interaction gain function. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0064] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0065] Figure 1 It is a schematic flowchart of a method for intelligent diagnosis of material damage by magnetic characteristic data fusion in an embodiment of the present invention. As Figure 1 shown, the method includes:

[0066] Based on the weld area of a pressure vessel, a magnetic stress detection robot equipped with a magnetic coercivity detection probe is used to collect magnetic characteristic data, including magnetic coercivity data, remanent magnetization intensity data, and magnetic permeability data;

[0067] Perform multi-scale decomposition on the magnetic characteristic data, and construct a magnetic domain evolution feature spectrum through the frequency band coefficient with the maximum energy entropy to obtain a magnetic domain wall migration feature map;

[0068] Based on the magnetic domain wall migration feature map, input the magnetic domain evolution feature spectrum into a composite neuron network to extract the magnetic domain structure feature, grain orientation feature, and texture evolution feature of the material, and obtain a material damage feature matrix;

[0069] In the material damage feature matrix, use tensor dimensionality reduction transformation and the hierarchical training mechanism of a multi-layer belief transfer network to construct a magnetic characteristic fusion feature vector;

[0070] Based on the magnetic characteristic fusion feature vector, realize fatigue-creep coupling damage calculation through temperature field modulation and adaptive weight allocation, and combine probability density analysis and an adaptive kernel function to obtain the material damage level and the remaining life value.

[0071] In an optional implementation manner, performing multi-scale decomposition on the magnetic characteristic data, and constructing a magnetic domain evolution feature spectrum through the frequency band coefficient with the maximum energy entropy to obtain a magnetic domain wall migration feature map includes:

[0072] Obtain a frequency band coefficient matrix by performing multi-layer wavelet packet decomposition on the collected magnetic characteristic data, and the frequency band coefficient matrix represents multi-scale characteristic information during the magnetization process;

[0073] Calculate the energy entropy value of the frequency band coefficient matrix to construct an energy distribution characteristic spectrum, identify the main frequency band magnetization response in the internal stress concentration region of the material based on the energy distribution characteristic spectrum, and select the main frequency band magnetization response as the characteristic frequency band;

[0074] Reconstruct the magnetization curve according to the time-frequency evolution law of the characteristic frequency band, perform magnetic domain wall pinning-unpinning motion analysis on the magnetization curve, and obtain the magnetic domain flipping kinetic density function through spectral analysis;

[0075] Construct a magnetic domain order parameter distribution matrix based on the main frequency band component of the magnetic domain flipping kinetic density function, which includes the anisotropic characteristics of damage evolution;

[0076] Perform coupled calculation on the magnetic domain order parameter distribution matrix and the internal dislocation density distribution of the material to obtain a two-dimensional characteristic function, and generate a magnetic domain wall migration characteristic map based on the two-dimensional characteristic function.

[0077] In a specific implementation, perform multi-layer wavelet packet decomposition on the collected magnetic characteristic data to obtain a frequency band coefficient matrix. The magnetic characteristic data collection can be obtained through a magnetic sensor array, the sampling frequency is set to 10 kHz, and 5000 data points are collected at each measurement point. Apply 5-layer wavelet packet decomposition to the collected magnetic characteristic data, and select the db4 wavelet basis function for decomposition. The decomposition process divides the signal into 32 different frequency bands, and each frequency band contains magnetization information within a specific frequency range. For example, the first frequency band range is 0 - 156.25 Hz, the second frequency band is 156.25 - 312.5 Hz, and so on. The frequency band coefficient matrix C(i, j) obtained through wavelet packet decomposition characterizes the multi-scale characteristic information in the magnetization process, where i represents the frequency band index (i = 1, 2,..., 32), and j represents the time index (j = 1, 2,..., 5000).

[0078] Calculate the energy entropy value of the frequency band coefficient matrix to construct an energy distribution characteristic spectrum. For each frequency band i, calculate its energy value Ei = ∑j|C(i, j)|² to obtain an energy distribution vector E = [E1, E2,..., E32] containing 32 elements. Normalize the energy distribution vector and calculate the energy entropy value H = -∑i(Ei / Etotal)×log2(Ei / Etotal), where Etotal = ∑iEi represents the total energy. In actual cases, the energy entropy value of healthy materials is about 3.85, while the energy entropy value of damaged materials is reduced to about 3.25. Identify the main frequency band magnetization response in the internal stress concentration region of the material based on the energy distribution characteristic spectrum, and select the frequency bands with an energy contribution rate exceeding 10% as the characteristic frequency bands. In the measured data, the 4th frequency band (468.75 - 625 Hz) and the 7th frequency band (937.5 - 1093.75 Hz) usually show as the main frequency bands, and the energy contribution rates reach 15.3% and 12.8% respectively.

[0079] Reconstruct the magnetization curve according to the time-frequency evolution law of the characteristic frequency band. Extract the characteristic frequency band coefficients C(4, j) and C(7, j), and apply the inverse wavelet packet transform to reconstruct the time-domain signal. Compared with the original signal, the reconstructed magnetization curve retains the main magnetization information, filters out the noise interference at the same time, and the signal-to-noise ratio is increased by about 5.8 dB. Analyze the magnetic domain wall pinning-unpinning motion of the reconstructed magnetization curve, obtain the instantaneous frequency of the magnetization signal through Hilbert transform, and construct a frequency-time distribution map. In the magnetic domain wall unpinning interval, the instantaneous frequency shows an obvious jump, the amplitude change exceeds 200 Hz, and the duration is about 0.5 ms. Obtain the magnetic domain flipping kinetic density function D(f) through spectral analysis, which characterizes the contribution of different frequency components to magnetic domain flipping. The kinetic density function is obtained by estimating the power spectral density of the reconstructed magnetization curve, and the frequency resolution is set to 2 Hz.

[0080] Construct a magnetic domain order parameter distribution matrix based on the main frequency band components of the magnetic domain flipping kinetic density function. Extract the frequency band range with the most concentrated energy in the kinetic density function, which is usually in the interval of 500 - 600 Hz and 950 - 1050 Hz. These frequency bands correspond to the main modes of magnetic domain flipping in the material. Construct a two-dimensional magnetic domain order parameter S(x, y) for each measurement point p(x, y), and its value is determined by the normalized energy of the main frequency band. In actual tests, the order parameter distribution in the healthy area is uniform, and the numerical fluctuation does not exceed ±0.15; while in the damaged area, the order parameter shows obvious abnormalities, the local change amplitude reaches ±0.45, and it shows obvious directional characteristics. This anisotropic characteristic is quantified by calculating the order parameter gradient matrix where the gradient magnitude and direction reflect the damage degree and orientation respectively.

[0081] Couple the magnetic domain order parameter distribution matrix with the dislocation density distribution inside the material to obtain a two-dimensional characteristic function. The dislocation density ρ(x, y) is estimated through the observation of the material microstructure, and the typical value is in the range of 10 10 -10 12 m -2 The two-dimensional characteristic function F(x, y) is calculated through weighted fusion: F(x, y) = w1×S(x, y) + w2×log10(ρ(x, y)), where the weight coefficients w1 = 0.65 and w2 = 0.35 are determined by fitting a large amount of experimental data. The value range of F(x, y) is usually between [-1, 1], where the negative value area corresponds to the low stress area and the positive value area corresponds to the high stress area. The area where the spatial gradient of the characteristic function exceeds 0.3 / mm is marked as potential damage points.

[0082] Generate a magnetic domain wall migration feature map based on a two-dimensional characteristic function. The adaptive threshold segmentation algorithm is used to process the characteristic function F(x, y), and the threshold is set to the mean plus 1.5 times the standard deviation. Morphological processing, including opening and closing operations, is applied to the segmented binary image, and a 3×3 matrix is selected as the structural element. The generated feature map enhances the visual effect through pseudo-coloring, with the damaged area marked in red, the healthy area marked in green, and the transition area marked in yellow.

[0083] In this embodiment, the band coefficient matrix obtained by multi-layer wavelet packet decomposition can characterize the fine multi-scale features in the magnetization process, improving the resolution ability of the microscopic magnetic response of the material; the energy distribution feature spectrum constructed based on the energy entropy value can accurately locate and extract the main band magnetization response in the stress concentration area inside the material, providing a more targeted characteristic band for subsequent analysis; the magnetization curve is reconstructed using the time-frequency evolution law of the characteristic band, and the kinematic analysis of the magnetic domain wall pinning-unpinning process is carried out, enabling the dynamic description of the dynamic behavior of the material during the loading process; the magnetic domain flipping kinetic density function obtained through spectrum analysis realizes the statistics and quantification of the flipping process, providing more physically meaningful parameters for the damage degree assessment; the magnetic domain order parameter distribution matrix constructed based on the main band components can reflect the direction-dependent characteristics during the damage evolution process, improving the recognition accuracy of the damage mode and direction; coupling the magnetic domain order parameter with the dislocation density distribution, the obtained two-dimensional characteristic function and the finally generated magnetic domain wall migration feature map can intuitively display the evolution of the internal microstructure and damage distribution of the material, providing a powerful tool for non-destructive testing and structural health monitoring.

[0084] In an alternative embodiment, for the magnetic domain wall pinning-unpinning motion analysis of the magnetization curve, the magnetic domain flipping kinetic density function obtained through spectrum analysis includes:

[0085] Calculate the local magnetic susceptibility of the reconstructed magnetization curve, and based on the comparison result between the local magnetic susceptibility and the preset critical magnetic susceptibility, divide the magnetization curve into a pinning-dominated region and a depinning-dominated region;

[0086] Establish a multi-pinning point potential field distribution function in the pinning-dominated region, including the pinning potential field strength coefficient, the pinning point position coordinates, and the pinning potential field action range parameter; construct a magnetic domain wall migration velocity field function in the depinning-dominated region, including the characteristic velocity parameter, the local effective magnetic field, and the critical magnetic field strength;

[0087] Substitute the multi-pinning point potential field distribution function and the magnetic domain wall migration velocity field function into the pinning-unpinning coupling equation, and solve to obtain the magnetic domain wall displacement response function, including the amplitude decay coefficient, the natural oscillation frequency, and the initial phase;

[0088] The autocorrelation operation is performed on the magnetic domain wall displacement response function and the Fourier transform is performed to obtain the kinetic density function, which characterizes the transition characteristics of the magnetic domain wall between pinning points.

[0089] In a specific embodiment, the local magnetic susceptibility is calculated for the reconstructed magnetization curve, which is achieved by differentiating between adjacent points on the magnetization curve. For each sampling point i, the local magnetic susceptibility χi=(Mi + 1 - Mi) / (Hi + 1 - Hi) is calculated, where M represents the magnetization intensity and H represents the applied magnetic field intensity. The sampling interval is set to 0.05 microseconds to ensure capturing the transient changes during the magnetization process. The calculated local magnetic susceptibility curve exhibits non-linear characteristics. The maximum magnetic susceptibility value of the healthy material is approximately 85 A / m·T, and that of the damaged material drops to 65 A / m·T. The preset critical magnetic susceptibility χc is 40% of the maximum magnetic susceptibility, i.e., 34 A / m·T for the healthy material and 26 A / m·T for the damaged material. Based on the comparison results of the local magnetic susceptibility and the critical magnetic susceptibility, when χi < χc, the corresponding region is divided into the pinning-dominated region; when χi ≥ χc, the corresponding region is divided into the depinning-dominated region. In practical applications, the magnetization curve of the 45 steel specimen is divided into 4 pinning-dominated regions and 3 depinning-dominated regions, and the boundary points appear at the applied magnetic field intensities of 156 A / m, 487 A / m, 892 A / m, 1268 A / m, 1624 A / m, and 1921 A / m.

[0090] The multi-pinning point potential field distribution function is established in the pinning-dominated region. The pinning potential field distribution function is expressed as V(r), which describes the constraint effect of the pinning points on the magnetic domain wall, where r represents the spatial position vector. V(r) consists of three key parameters: the pinning potential field strength coefficient α, the pinning point position coordinate ri, and the pinning potential field action range parameter β. In practical applications, for the 45 steel containing 0.45% carbon, the value range of the pinning potential field strength coefficient α is from 4.5×10^-6 to 6.8×10^-6 J / m², and different defect types correspond to different values; the pinning point position coordinate ri is determined through the analysis of the material microstructure. In this embodiment, 5 main pinning points are set, and the coordinates are r1=(2.3, 1.5)μm, r2=(4.7, 3.2)μm, r3=(5.6, 8.4)μm, r4=(8.9, 6.1)μm, r5=(11.2, 4.3)μm; the pinning potential field action range parameter β is related to the characteristic size of the pinning points, taking the value of 2.2 for carbide precipitates, 1.8 for dislocation clusters, and 2.5 for grain boundaries.

[0091] Construct a magnetic domain wall migration velocity field function in the depinning-dominated region. The velocity field function v(H) describes the motion velocity of the magnetic domain wall under the action of the effective magnetic field, and includes three key parameters: the characteristic velocity parameter v0, the local effective magnetic field Heff, and the critical magnetic field strength Hc. The characteristic velocity parameter v0 is related to the resistivity and saturation magnetization of the material, and takes a value of 18.5 m / s for 45 steel; the local effective magnetic field Heff takes into account the combined effects of the applied magnetic field, the demagnetizing field, and the pinning field, and is obtained through iterative calculation, and its numerical range is between 85% and 110% of the applied magnetic field; the critical magnetic field strength Hc represents the threshold at which the magnetic domain wall begins to move on a large scale, and takes a value of 123 A / m in this embodiment. In the damaged material, due to the uneven distribution of internal stress, the critical magnetic field strength exhibits a spatial distribution characteristic, and can increase to 225 A / m in the crack tip region. The calculation of the velocity field function adopts a segmented processing method. When Heff < Hc, v(H) is close to zero; when Heff ≥ Hc, v(H) approximately increases linearly with a slope of 0.15 m·m / A·s.

[0092] Substitute the multi-pinning point potential field distribution function and the magnetic domain wall migration velocity field function into the pinning-depinning coupling equation to solve for the magnetic domain wall displacement response function. The coupling equation is expressed in terms of three parameters: displacement, velocity, and acceleration, and is solved by a numerical integration method. The integration step size is set to 0.01 microseconds, and the fourth-order Runge-Kutta algorithm is used to ensure the calculation accuracy. The obtained magnetic domain wall displacement response function x(t) includes the amplitude decay coefficient λ, the natural oscillation frequency f0, and the initial phase Three characteristic parameters. For healthy materials, the amplitude decay coefficient λ is approximately 850 s^-1, the natural oscillation frequency f0 is approximately 5.6 kHz, and the initial phase is close to π / 6; for damaged materials containing microcracks, the amplitude decay coefficient increases to 1250 s -1 , the natural oscillation frequency decreases to 4.2 kHz, and the initial phase becomes about π / 4. The average amplitude of the displacement response function reflects the stability of the magnetic domain wall motion. For healthy materials, it is about 1.8 μm, and for damaged materials, it decreases to 1.2 μm and is accompanied by obvious irregular fluctuations.

[0093] The autocorrelation operation is performed on the magnetic domain wall displacement response function and the dynamic density function is obtained through Fourier transform. The autocorrelation operation is implemented by the sliding window method with a window length of 2048 data points and an overlap rate of 50%. The fast Fourier transform is applied to the autocorrelation function with a sampling frequency of 100 kHz to obtain the dynamic density function S(f) with a frequency resolution of 48.8 Hz. The dynamic density function characterizes the transition characteristics of the magnetic domain wall between pinning points, where the main frequency peak corresponds to the characteristic frequency of magnetic domain flipping, and the peak width reflects the consistency of the magnetic domain wall movement. In actual tests, the dynamic density function of healthy No. 45 steel presents a single sharp peak at 5.6 kHz, and the full width at half maximum is about 280 Hz; the main frequency peak of the micro-damaged material shifts down to 4.2 kHz, the full width at half maximum increases to 450 Hz, and secondary peaks appear at 2.8 kHz and 7.5 kHz. Statistical features such as the skewness and kurtosis of the dynamic density function can quantify the degree of damage. The skewness of the healthy material is about 0.12, and the kurtosis is about 3.05; while the skewness of the damaged material increases to 0.38 and the kurtosis decreases to 2.65, indicating that the magnetic domain flipping process becomes irregular.

[0094] Traditional magnetic domain wall motion analysis techniques mainly rely on Barkhausen noise measurement to evaluate damage by statistically analyzing the amplitude and frequency characteristics of the noise, but they cannot accurately distinguish the pinning process and the depinning process, resulting in limited diagnostic accuracy in materials with complex microstructures. In the prior art, the interaction between the magnetic domain wall and the pinning points is simplified, and usually a single pinning potential field model is adopted, which is difficult to describe the cooperative effect of multiple types of pinning points in actual materials. The solution of this embodiment realizes the accurate division of the magnetization curve in the pinning-dominated region and the depinning-dominated region by introducing local susceptibility analysis; establishes a multi-pinning point potential field distribution function, breaking through the limitations of the traditional single potential field model; constructs a magnetic domain wall migration velocity field function considering the local effective magnetic field distribution, improving the adaptability to inhomogeneous materials; the magnetic domain wall displacement response function obtained by solving the pinning-depinning coupling equation truly reflects the modulation effect of the material microstructure on the magnetization process; the microscopic magnetic changes caused by damage are accurately captured through the spectral characteristics of the dynamic density function.

[0095] As Figure 2 shown, by calculating the local susceptibility (orange curve) of the reconstructed magnetization curve (blue solid line) and comparing it with the critical susceptibility value χc = 4.82×10 -3(The red dashed line) is compared, and the magnetization process is successfully divided into a pinning-dominated region (the purple region) and a depinning-dominated region (the yellow region). The figure shows that in the range of the applied magnetic field H = 1850 - 3650 A / m, the local magnetic susceptibility value is lower than the critical value, corresponding to the pinning-dominated mechanism; when H > 3650 A / m, the local magnetic susceptibility rises sharply above the critical value, indicating that the material enters the depinning-dominated region. Three key characteristic points are specifically marked in the figure: the red dot represents the critical point (H = 3650 A / m, M = 1.25×10 5 A / m) of the transformation of the magnetic domain wall motion mechanism, at which time the local magnetic susceptibility is exactly equal to the critical value χc = 4.82×10 -3 ; the green dot marks the point of the minimum local magnetic susceptibility (H = 2500 A / m, χ = 2.35×10 -3 ), indicating that the pinning effect inside the material is the strongest at this magnetic field intensity, and this characteristic may reveal the region with the highest pinning point density in the microstructure of the material; the purple dot identifies the point of the maximum local magnetic susceptibility (H = 4000 A / m, χ = 8.75×10 -3 ), at which time the depinning motion of the magnetic domain wall is the most active, corresponding to the peak value of the magnetic domain wall migration speed. Compared with the traditional Barkhausen noise analysis technology, this technical solution can not only clearly identify the transformation point of the magnetic domain wall motion mechanism, but also capture the local magnetic susceptibility fluctuations caused by the inhomogeneity of the material microstructure, which is of great significance for understanding the cooperative effect of multiple types of pinning points in actual materials. Through this precise regional division, it lays a data foundation for the subsequent establishment of the potential field distribution function of multiple pinning points and the magnetic domain wall migration speed field function, and significantly improves the accuracy of the microstructure characterization of magnetic materials.

[0096] In an alternative embodiment, based on the magnetic domain wall migration feature map, the magnetic domain evolution feature spectrum is input into a composite neural network to extract the magnetic domain structure features, grain orientation features, and texture evolution features of the material, and a material damage feature matrix is obtained, including:

[0097] The magnetic domain evolution feature spectrum is input into a composite neural network including a parallel convolutional layer. The parallel convolutional layer is provided with convolutional kernels of different scales. The characteristic components of the magnetic domain wall migration are separated through the parallel convolutional layer, and a magnetic domain structure feature vector is obtained through a feature enhancement layer;

[0098] A non-linear transformation is performed on the magnetic domain evolution feature spectrum in the feature mapping layer of the composite neural network to extract the spatial distribution information of the grain arrangement and generate a grain orientation feature vector;

[0099] The texture information in the magnetic domain evolution feature spectrum is separated through the feature decoupling layer of the composite neural network, and a texture evolution feature vector is obtained by using a dynamic feature extraction unit;

[0100] According to the information entropy criterion, weight coefficients are assigned to the magnetic domain structure feature vector, the grain orientation feature vector, and the texture evolution feature vector, and a feature interaction matrix is established to calculate the correlation degree, obtaining a multi-feature fusion vector.

[0101] The multi-feature fusion vector is input into the damage feature mapping network, and feature transformation is performed through the residual connection structure and the non-linear activation function, outputting the material damage feature matrix.

[0102] In a specific embodiment, the magnetic domain evolution feature spectrum is input into a composite neuron network including a parallel convolutional layer for feature extraction processing. The magnetic domain evolution feature spectrum is obtained through the magneto-optical Kerr effect microscopy imaging technology, with a collection resolution of 1024×1024 pixels and a time resolution of 100 frames per second, recording the magnetic domain structure changes during the process of the applied magnetic field increasing from 0 to 2000 A / m. The composite neuron network adopts a multi-branch structure, and the parallel convolutional layer is set with three different scales of convolutional kernels, which are 3×3, 5×5, and 7×7 pixels respectively, and the number of convolutional kernels is 64 each. The small-size convolutional kernel (3×3) mainly extracts the fine change features of the magnetic domain walls; the medium-size convolutional kernel (5×5) captures the cooperative motion patterns of the magnetic domain groups; the large-size convolutional kernel (7×7) identifies the overall change trend of the large-scale magnetic domain structure. The stride of each convolutional kernel is set to 2 pixels, and the padding method adopts zero padding to keep the size of the feature map. After separating the feature components of the magnetic domain wall migration through the parallel convolutional layer, the feature enhancement layer uses the channel attention mechanism to enhance the key features, and the channel weight coefficients are calculated through global average pooling and a two-layer fully connected network. In practical applications, for the 45 steel material, the distribution range of the weight coefficients of the magnetic domain wall features in the channel dimension is from 0.65 to 1.35, and the weight coefficients of the features near the magnetic domain wall pinning points are generally higher than 1.2. The dimension of the magnetic domain structure feature vector obtained after being processed by the feature enhancement layer is 512, including key information such as the magnetic domain morphology, the magnetic domain wall density, and the magnetic domain wall movement speed.

[0103] Perform a non - linear transformation on the magnetic domain evolution feature spectrum in the feature mapping layer of the composite neuron network to extract the spatial distribution information of grain arrangement. The feature mapping layer consists of three sub - modules: a spatial transformation module, a direction - sensitive module, and a scale transformation module. The spatial transformation module uses a deformable convolutional network, sets a 9×9 convolutional kernel, and an offset learning rate of 0.005, and can adaptively adjust the receptive field shape to match the irregular grain boundaries. The direction - sensitive module contains 8 direction filters with an angular interval of 22.5 degrees, and the output channel number of each direction filter is 32, effectively extracting the directional features of grain arrangement. The scale transformation module adopts a dilated convolutional structure with dilation rates of 1, 2, and 4 respectively, comprehensively capturing grain features at different scales. For the 45 - steel material, the average grain size is about 25 microns. After being processed by the feature mapping layer, the system can identify grain boundaries in the range of 20 - 30 microns with an accuracy of ±2 microns. The dimension of the grain orientation feature vector output by the feature mapping layer is 384, containing information such as grain size distribution, grain boundary density, and grain orientation distribution. In the fatigue damage samples, the grain orientation feature vector shows that the characteristic response value at the grain boundaries is 45% to 70% higher than that inside the grains, while this difference is only 25% to 40% in healthy materials.

[0104] Separate the texture information in the magnetic domain evolution feature spectrum through the feature decoupling layer of the composite neuron network. The feature decoupling layer adopts an attention - guided auto - encoder structure, which consists of an encoder and a decoder. The encoder is composed of three layers of residual convolutional blocks, and each layer of residual convolutional block contains two 3×3 convolutional layers and a skip connection. The activation function uses LeakyReLU with a negative slope of 0.2. The decoder adopts a transposed convolutional structure with an up - sampling rate of 2, and guides the network to learn the separated representation of texture features through reconstruction loss and adversarial loss. During the feature decoupling process, the channel attention threshold is set to 0.4, and the channel features below this threshold are suppressed, retaining the high - response channels related to texture. Use a dynamic feature extraction unit to obtain the texture evolution feature vector. The dynamic feature extraction unit contains a long short - term memory network structure, with 256 hidden layer nodes, a time step set to 10 frames, and a forget - gate threshold of 0.25. For the 45 - steel material, the dynamic feature extraction unit can capture the texture evolution process caused by the change of the applied magnetic field. Especially in the range of magnetic field strength from 450 A / m to 850 A / m, the time derivative value of the texture evolution feature vector increases significantly, reaching 1.8 to 2.5 times that of the healthy state, indicating an intensification of the internal microstructure change of the material. The finally obtained texture evolution feature vector has a dimension of 256, containing key information such as texture type, texture strength, and texture evolution rate.

[0105] Weight coefficients are assigned to the magnetic domain structure feature vector, grain orientation feature vector, and texture evolution feature vector according to the information entropy criterion. During the information entropy calculation process, each feature vector is normalized and its probability distribution is calculated. The sliding window method is used to calculate the local information entropy, with the window size being 10% of the feature vector dimension and the sliding step being 5. The higher the information entropy value of the feature vector, the richer the effective information it contains, and the larger the corresponding weight coefficient. In practical applications, the information entropy of the magnetic domain structure feature vector of 45 steel material is 5.85, and the weight coefficient is 0.42; the information entropy of the grain orientation feature vector is 4.92, and the weight coefficient is 0.35; the information entropy of the texture evolution feature vector is 3.23, and the weight coefficient is 0.23. A feature interaction matrix is established to calculate the correlation degree. The dimension of the feature interaction matrix is (512 + 384 + 256) × (512 + 384 + 256), and the correlation strength between features is calculated through cosine similarity. The correlation strength threshold is set to 0.65, and feature pairs exceeding the threshold are regarded as strongly correlated and given additional enhancement during the multi-feature fusion process. For damaged materials, the correlation strength between the magnetic domain wall feature and the grain boundary feature is significantly higher than that of healthy materials, with an increase ranging from 35% to 45%. This correlation enhancement phenomenon is an early indicator of microscopic damage. The multi-feature fusion vector is generated through weighted summation and a non-linear activation function, with a final dimension of 512, containing comprehensive information on the material's magnetism, microstructure, and damage state.

[0106] The multi - feature fusion vector is input into the damage feature mapping network, and feature transformation is performed through the residual connection structure and the non - linear activation function. The damage feature mapping network contains 5 residual blocks. Each residual block contains two convolutional layers and a skip connection. The convolutional kernel size is 3×3, and the number of channels is 512. The residual connection structure effectively alleviates the vanishing gradient problem of the deep network and ensures the information flow during the feature transformation process. The non - linear activation function uses the Swish function, which provides a smoother gradient and better feature expression ability than the traditional ReLU function. The learning rate of the damage feature mapping network is set to 0.0003 and dynamically adjusted using the cosine annealing strategy, with the minimum learning rate being 5% of the initial value. The network training uses the mini - batch gradient descent method with a batch size of 64, and the number of training samples is 5000, including material data with different damage degrees. The dimension of the material damage feature matrix output by the damage feature mapping network is 32×32, and the value range of each element is from 0 to 1, representing the damage probability at the corresponding position. In the fatigue damage sample test of 45# steel, the damage feature matrix shows a high damage probability value above 0.85 in the crack tip region, a damage probability value below 0.15 in the healthy region, and the damage probability distribution in the boundary transition region conforms to the stress concentration attenuation law. After the damage feature matrix is restored to the original image size (1024×1024) through bicubic interpolation, the coincidence degree with the actual damage position observed by the optical microscope reaches 92.8%, which improves the accuracy by more than 25% compared with the traditional magnetic detection method, and can detect subsurface damage with a depth of up to 3 mm.

[0107] In this embodiment, the parallel convolutional layer uses convolutional kernels of different scales, which can accurately separate the multi - scale feature components of the magnetic domain wall migration, improving the ability to capture subtle migration behaviors; the feature enhancement layer weights and strengthens the separated feature components to generate a high - dimensional magnetic domain structure feature vector, significantly enhancing the discrimination of the structure representation; the feature mapping layer extracts the spatial distribution information of the grain arrangement through non - linear transformation to construct a grain orientation feature vector, providing a microstructural basis for subsequent damage localization; the feature decoupling layer combines dynamic feature extraction units to independently extract and vectorize the texture evolution features, effectively separating the contribution of texture changes to the overall damage; based on the information entropy criterion, the three types of feature vectors are weighted and fused, and non - linear transformation is performed through the damage feature mapping network with residual connection, and finally a high - precision material damage feature matrix is output.

[0108] In an alternative embodiment, in the material damage feature matrix, using the tensor dimensionality reduction transformation and the hierarchical training mechanism of the multi - layer belief propagation network, constructing the magnetic property fusion feature vector includes:

[0109] The material damage feature matrix is reconstructed into a third - order tensor through the tensor outer product operation, and the third - order tensor is determined by the tensor decomposition factor and the tensor rank to form the damage feature tensor;

[0110] Perform Tucker decomposition on the damage feature tensor, construct the tensor product of the core tensor and the projection matrices of each modality, and obtain the dimensionality-reduced feature tensor;

[0111] Establish node state vectors in the multi-layer belief propagation network, calculate the belief propagation values between nodes based on the connection weights, and form the network belief state matrix;

[0112] Input the dimensionality-reduced feature tensor into the multi-layer belief propagation network, construct a hierarchical loss function based on the reconstruction loss function, belief propagation loss function, and consistency constraint loss function, and perform hierarchical training;

[0113] According to the gradient information of the hierarchical loss function, use the learning rate parameter and regularization parameter to iteratively update the network belief state matrix, generate the feature weight matrix, and extract global features and local features based on the feature weight matrix;

[0114] Based on the feature weight matrix, perform weighted combination on the global feature and the local feature, and obtain the magnetic property fusion feature vector through the activation function.

[0115] In a specific embodiment, the material damage feature matrix is reconstructed into a third-order tensor through tensor outer product operation. The size of the material damage feature matrix is 32×32, and each element represents the damage probability value at the corresponding position. During the tensor outer product operation, the feature matrix is unfolded along three dimensions to construct a three-dimensional representation of space-frequency-time. In actual operation, the feature matrix is transformed, frequency dimension information is introduced, and 5 key frequency band features (0 - 200Hz, 200 - 500Hz, 500 - 1000Hz, 1000 - 2000Hz, 2000 - 5000Hz) are extracted using the short-time Fourier transform, and at the same time, the sequence information of 10 time frames is retained. The size of the third-order tensor obtained after the tensor outer product operation is 32×5×10, that is, the spatial dimension is 32, the frequency dimension is 5, and the time dimension is 10. This third-order tensor is determined by the tensor decomposition factor and the tensor rank to form the damage feature tensor. In this embodiment, the tensor decomposition factor is set to 8, indicating the number of principal components retained in each dimension; the tensor rank is set to 16, representing the complexity level of the feature tensor. For the fatigue damage samples of 45 steel, the characteristic response value of the low-frequency band (0 - 200Hz) in the damaged area is 65% higher than that in the healthy area, while the response of the high-frequency band (2000 - 5000Hz) in the damaged area is about 40% lower than that in the healthy area. This spectral characteristic difference is an important basis for damage identification.

[0116] Perform Tucker decomposition on the damage feature tensor, construct the tensor product of the core tensor and the projection matrices of each modality, and obtain the dimension-reduced feature tensor. Tucker decomposition decomposes the original feature tensor into a combination of a core tensor and three projection matrices. In this implementation, the size of the core tensor is set to 8×4×6, and the three projection matrices are a 32×8-dimensional spatial modality matrix, a 5×4-dimensional frequency modality matrix, and a 10×6-dimensional time modality matrix respectively. The alternating least squares method is used in the decomposition process, the upper limit of the number of iterations is set to 200, and the convergence threshold is 1e-6. To improve the decomposition efficiency, a random initialization strategy is adopted, and the initialization range is a uniform distribution in [-0.1, 0.1]. In the test data of 45 steel fatigue, the first three principal components of the spatial modality matrix explain 85.7% of the spatial variance, the first two principal components of the frequency modality matrix explain 92.3% of the frequency variance, and the first four principal components of the time modality matrix explain 88.9% of the time variance. The dimension-reduced feature tensor obtained through the tensor product operation has a size of 8×4×6, and the data volume is reduced by approximately 90%, while retaining more than 95% of the information of the original tensor. In the damage samples, the sparsity (proportion of zero elements) of the core tensor is approximately 75%, which is significantly higher than 62% of the healthy samples, indicating that the features are more concentrated in the damage state.

[0117] Establish node state vectors in the multi-layer belief propagation network, calculate the belief propagation values between nodes based on the connection weights, and form the network belief state matrix. The multi-layer belief propagation network consists of 4 layers, and the number of nodes in each layer is 192, 128, 64, and 32 respectively. Each node maintains a state vector, and the dimension of the initial state vector is the same as the total number of elements of the dimension-reduced feature tensor, which is 8×4×6 = 192. The connections between nodes in the network are represented by a graph structure, and the connection weights between adjacent layers are initialized by a pre-trained autoencoder, and the weight range is controlled within [-0.2, 0.2]. The belief propagation mechanism between nodes is implemented based on the message passing algorithm. In each iteration, the node state update depends on the information passed by adjacent nodes and the node's own state. In this embodiment, the belief propagation value is calculated by the inner product of the node state vector and the connection weight, and is normalized by the Sigmoid function, with a value range of [0, 1]. The number of belief propagation iterations is set to 15, and the iteration convergence threshold is 0.01. For the 45 steel fatigue samples, the node state update rate in the damage area reaches 1.8 times that of the healthy area, indicating that the damage features are more active in the belief network. The size of the network belief state matrix is the total number of network nodes × the dimension of the state vector, that is, (192 + 128 + 64 + 32)×192 = 79872 elements, and each row represents the belief state of a node.

[0118] The dimensionality-reduced feature tensor is input into a multi-layer belief propagation network, and a hierarchical loss function is constructed based on a reconstruction loss function, a belief propagation loss function, and a consistency constraint loss function for hierarchical training. The dimensionality-reduced feature tensor is first flattened into a 192-dimensional vector as the input to the nodes in the first layer of the network. The reconstruction loss function uses the mean squared error to calculate the difference between the network output and the original feature tensor, and the weight coefficient is set to 0.5. The belief propagation loss function measures the consistency of the belief states between adjacent nodes, adopts the form of cross-entropy, and the weight coefficient is 0.3. The consistency constraint loss function constrains the network parameters through a regularization term, adopts the form of L1 norm, and the weight coefficient is 0.2. The hierarchical loss function combines the three losses at different levels. The lower-layer network focuses on the reconstruction loss, the middle layer balances the three losses, and the upper-layer network focuses on the belief propagation loss and the consistency constraint. The hierarchical training strategy adopts a layer-by-layer optimization method. Each layer is trained for 500 rounds, the batch size is 32, the initial learning rate is 0.001, and it decays to 0.8 times the original every 100 rounds. For the test samples, the value of the hierarchical loss function drops from the initial 3.85 to 0.42 at convergence, where the reconstruction loss drops from 2.12 to 0.18, the belief propagation loss drops from 1.05 to 0.16, and the consistency constraint loss drops from 0.68 to 0.08, indicating that the network has successfully learned the deep feature representation of material damage.

[0119] According to the gradient information of the hierarchical loss function, the learning rate parameter and the regularization parameter are used to iteratively update the network belief state matrix to generate a feature weight matrix. Based on the feature weight matrix, global features and local features are extracted. The gradient information is calculated by the backpropagation algorithm, and the gradient clipping threshold is set to 5.0 to prevent gradient explosion. The learning rate parameter adopts an adaptive adjustment strategy, with an initial value of 0.001. When the change in the loss function is less than 0.001 for 5 consecutive rounds, the learning rate is halved. The regularization parameter is set to 0.0005 to control the model complexity. After 30 rounds of iterative update of the network belief state matrix, a feature weight matrix is generated, with a size of 192×32, indicating the contribution weights of 192 original features to 32 abstract features. In the experiment, the sparsity of the feature weight matrix is about 82%, indicating that each abstract feature is dominated by only a few original features. Global features and local features are extracted based on the feature weight matrix. The global features are obtained by performing SVD decomposition on the feature weight matrix to get the principal components, and the first 8 principal components are taken to form a 64-dimensional global feature vector; the local features are extracted from the feature weight matrix by the local sensitive hashing method to obtain 32 local patterns, which are combined to form a 128-dimensional local feature vector. In the test samples of 45# steel, the recognition accuracy of the global features for the overall damage degree of the material reaches 94.2%, and the positioning accuracy of the local features for the damage location reaches ±1.5mm.

[0120] Based on the feature weight matrix, the global features and local features are weighted and combined, and after passing through the activation function, a magnetic property fusion feature vector is obtained. During the weighted combination process, the global feature weight coefficient is 0.4, and the local feature weight coefficient is 0.6, reflecting the importance of local damage features in material damage diagnosis. The feature combination adopts a non-linear mapping method and is realized through a fully connected network. The number of hidden layer nodes is 96, and the GELU function is used as the activation function. The GELU function has better gradient characteristics than the traditional ReLU function and provides a smoother non-linear transformation. The dimension of the magnetic property fusion feature vector obtained after passing through the activation function is 128, which contains comprehensive information about material damage. In the fatigue test of 45 steel, the first 10 components of the fusion feature vector correspond to damage type information. Component values in the range of 0.75 - 0.85 indicate fatigue crack damage, and in the range of 0.55 - 0.65 indicate corrosion damage; the middle 50 components correspond to damage location and distribution information, and the damage area and healthy area are distinguished by a threshold of 0.6; the last 68 components correspond to damage degree information, which is negatively correlated with the remaining life of the material, and the correlation coefficient reaches -0.87. The final magnetic property fusion feature vector realizes the comprehensive characterization of the material damage type, location, and degree, providing a reliable feature basis for subsequent damage assessment and life prediction. The consistency between the magnetic property fusion feature vector and the actual damage state reaches 91.5%, which improves the diagnostic accuracy by 23.6% compared with the single magnetic property feature, and the detection sensitivity for micro-damage is increased by 2.8 times. The minimum crack length that can be detected is reduced from 1.2 mm of the traditional method to 0.43 mm.

[0121] In this embodiment, the damage feature matrix is reconstructed into a third-order tensor through the outer product of tensors, and the tensor rank constraint is used to achieve a high-dimensional representation of complex damage information; the Tucker decomposition is used to decompose the third-order tensor into a core tensor and modal projection matrices, and the reduced-dimensional feature tensor is reconstructed through the tensor product, retaining key damage features while greatly reducing the dimension; node state vectors are established in the multi-layer belief propagation network and beliefs are propagated based on connection weights to effectively capture the correlation between different modalities and spatial positions, forming a network belief state matrix; a hierarchical training objective is constructed based on reconstruction, belief propagation, and consistency constraint loss, and the network belief state is iteratively updated using gradient information, learning rate, and regularization parameters to generate a refined feature weight matrix; the global and local features are extracted and weighted and combined through the feature weight matrix, and after non-linear activation, a fused magnetic property feature vector is obtained, providing efficient and robust feature support for accurate detection and quantification of material damage.

[0122] In an alternative embodiment, based on the magnetic property fusion feature vector, fatigue-creep coupling damage calculation is realized through temperature field modulation and adaptive weight allocation, and combined with probability density analysis and adaptive kernel function, the material damage grade and remaining life value are obtained, including:

[0123] Extract the magnetic parameter group from the fused magnetic feature vector, perform non-linear mapping on the magnetic parameter group through the temperature field modulation function, establish the correlation law between the magnetic field and the stress field, and calculate the fatigue damage variable;

[0124] Determine the creep stress correction coefficient by passing the fused magnetic feature vector through the adaptive weight assignment algorithm, and calculate the creep damage variable in combination with the creep strain evolution law;

[0125] Perform coupling processing on the fatigue damage variable and the creep damage variable, calculate the field effect superposition amount using the non-linear coupling function, establish the damage accumulation path, and obtain the total damage amount;

[0126] Calculate and analyze the distribution characteristics of the total damage amount using the piecewise probability density, calculate the two-parameter life distribution probability value using the adaptive kernel function, and realize the dynamic assessment of the material damage state;

[0127] Set the hierarchical damage assessment criteria according to the two-parameter life distribution probability value, determine the multi-level damage discrimination boundary in combination with the material damage critical threshold, and calculate the material damage grade;

[0128] Substitute the total damage amount and the material damage grade into the damage evolution rate equation, determine the life attenuation law through integral calculation, and output the remaining life value of the material.

[0129] In a specific embodiment, magnetic parameter groups in the magnetic property fusion feature vector are extracted, and a non-linear mapping is performed on the magnetic parameter groups through a temperature field modulation function. The magnetic parameter groups include key magnetic parameters such as coercive force, remanence, magnetic permeability, and hysteresis loss, which are extracted from the first 32 components of the 128-dimensional magnetic property fusion feature vector. For 45# steel material, the typical coercive force value ranges from 400 to 600 A / m, the remanence value ranges from 0.8 to 1.2 T, and the initial magnetic permeability is between 150 and 250. The temperature field modulation function is implemented in the form of a piecewise continuous function to achieve non-linear mapping, taking into account the influence law of temperature on magnetic parameters. The reference temperature in the modulation function is set to 25 °C, and the modulation coefficient varies within the range of 0.85 - 1.15. In practical applications, when the ambient temperature is 120 °C, the coercive force value decreases by about 18%, and the remanence value decreases by about 12%. These changes are corrected by the temperature coefficients of -0.0015 / °C and -0.001 / °C. The magnetic parameters after temperature field modulation are correlated with the internal stress field of the material, and the correlation law is stored in the form of a look-up table, containing 2000 discrete points, and the stress range covers 0 - 800 MPa. Based on the magneto-elastic coupling principle, when the material stress increases to 300 MPa, the coercive force increases by about 45 A / m, and the remanence decreases by about 0.15 T. The fatigue damage variable is calculated through this correlation law, and the fatigue damage variable ranges from 0 to 1, corresponding to the state of the material from no damage to complete failure. For a 45# steel sample subjected to 50,000 cyclic loads, the fatigue damage variable calculated from the magnetic parameters is 0.35, indicating that the material has consumed 35% of its fatigue life.

[0130] The creep stress correction coefficient is determined by passing the magnetic property fusion feature vector through an adaptive weight allocation algorithm. The adaptive weight allocation algorithm is based on the entropy weight method, calculates the information entropy of the 33rd to 64th components in the magnetic property fusion feature vector, and assigns weight coefficients according to the information entropy values. In the calculation of information entropy, 10 equal-width intervals are used to divide the feature distribution, and the maximum entropy value is set to 3.32. In the creep damage samples, the typical information entropy distribution is 2.45 - 2.85, and the corresponding weight coefficients are 0.12 - 0.25. The creep stress correction coefficient is calculated through weighted combination and is used to adjust the stress parameters in the standard creep model. For 45# steel at a working temperature of 550 °C and a stress level of 180 MPa, the calculated creep stress correction coefficient is 1.28, indicating that the actual effective stress is 28% higher than the nominal stress. The creep damage variable is calculated in combination with the creep strain evolution law. The creep strain evolution is described by a three-stage model, including the initial creep stage, the steady-state creep stage, and the accelerated creep stage. The creep rate is set to 3.2×10 -8 / s in the initial stage, 1.5×10 -8 / s in the steady state stage, and the maximum value in the accelerated stage can reach 8.7×10 -8 / s. The creep damage variable is calculated by the ratio of the cumulative creep strain to the fracture strain. When the cumulative strain reaches 70% of the fracture strain, the creep damage variable of the material is 0.7, indicating that 70% of the creep life has been consumed.

[0131] The fatigue damage variable and the creep damage variable are coupled, and the field effect superposition amount is calculated using a non-linear coupling function. The non-linear coupling function adopts the form of a modified hyperbolic tangent function, which includes two key parameters: the interaction coefficient γ and the shape parameter α. The interaction coefficient γ reflects the mutual promotion effect between fatigue and creep damage, and its value for No. 45 steel is 1.35; the shape parameter α controls the non-linearity degree of the coupling curve, and its value is 2.2. The field effect superposition amount represents the additional damage generated by the combined action of fatigue and creep, and its value range is 0 - 0.5. In a high-temperature and high-stress environment, when the fatigue damage variable is 0.35 and the creep damage variable is 0.28, the calculated field effect superposition amount is 0.17, indicating that the fatigue-creep interaction significantly increases the overall damage level. When establishing the damage accumulation path, 64 path nodes are used to record the damage evolution process. The node spacing is set to 0.01 in the initial stage of damage and gradually decreases to 0.005 in the later stage of damage to improve the characterization accuracy of the rapid damage stage. The final total damage amount is the combination of the fatigue damage variable, the creep damage variable, and the field effect superposition amount. For the above case, the total damage amount is 0.35 + 0.28 + 0.17 = 0.8, indicating that the material has consumed 80% of its overall life.

[0132] The distribution characteristics of the total damage amount are calculated and analyzed using piecewise probability density. The piecewise probability density calculation uses a 5-segment piecewise model, and each segment is described by a different probability density function. The connection points are set to 0.2, 0.4, 0.6, and 0.8 of the total damage amount. In the initial stage of damage (0 - 0.2), the Weibull distribution is used, with a shape parameter of 2.5; in the middle stage (0.2 - 0.6), the lognormal distribution is used, with a mean parameter of -0.5 and a standard deviation of 0.28; in the later stage (0.6 - 1.0), the generalized extreme value distribution is used, with a location parameter of 0.75, a scale parameter of 0.12, and a shape parameter of -0.35. The probability cumulative distribution curve calculated by the piecewise probability density function can accurately reflect the probability distribution characteristics of material damage. The probability value of the two-parameter life distribution is calculated using an adaptive kernel function. The initial value of the bandwidth of the adaptive kernel function is set to 0.15 and dynamically adjusted to the optimal value of 0.08 through the cross-validation method. The two-parameter life distribution includes a scale parameter η and a shape parameter β. For the fatigue-creep composite damage of No. 45 steel, the scale parameter η is 2.35×10 5 (number of cycles), and the shape parameter β is 3.25. The dynamic assessment of the material damage state is achieved by monitoring the change rate of the total damage amount. The change rate threshold is set to 0.05 / 1000 hours. Exceeding this threshold indicates that the damage has entered the acceleration stage, and the monitoring frequency needs to be increased.

[0133] Set up a hierarchical damage assessment criterion according to the probability value of the two-parameter lifetime distribution. The hierarchical damage assessment criterion divides the damage degree into 5 levels: micro-damage (0 - 0.3), mild damage (0.3 - 0.5), moderate damage (0.5 - 0.7), severe damage (0.7 - 0.9), and critical damage (0.9 - 1.0). Each damage level corresponds to different detection periods and maintenance strategies. Determine the multi-level damage discrimination boundary in combination with the critical threshold of material damage. The critical threshold is obtained through the statistics of a large amount of failure data. The fatigue critical threshold of 45 steel is 0.85, the creep critical threshold is 0.75, and the comprehensive critical threshold is 0.9. The multi-level damage discrimination boundary is calculated by the Bayesian decision surface, adopting the form of a quadratic discriminant function, and the dimension of the feature space is 3 (fatigue damage, creep damage, total damage). In the discrimination process, a cost function is introduced to adjust the decision boundary. The cost of misjudgment is set to 3.0, and the cost of false judgment is set to 1.0, reflecting that the safety risk caused by misjudgment is greater. When calculating the material damage level, the fuzzy membership method is used to process the samples in the boundary region, and the membership function adopts the Gaussian form with a standard deviation parameter of 0.05. For a sample with a total damage amount of 0.8, the calculated damage level is level 4 (severe damage), and the membership degree is 0.92, and maintenance is required within 200 hours.

[0134] Substitute the total damage amount and the material damage level into the damage evolution rate equation, and determine the law of life attenuation through integral calculation. The damage evolution rate equation adopts the modified Norton - Bailey form, which includes three key parameters: the non-linear exponent m of damage accumulation, the stress sensitivity exponent n, and the temperature dependence coefficient Q. For 45 steel, under the working condition of 550 °C, the value of m is 1.35, the value of n is 4.5, and the value of Q is 15000 K. The damage evolution rate accelerates with the increase of the total damage amount. When the total damage amount exceeds 0.7, the evolution rate increases by about 2.5 times. The integral calculation adopts the fourth-order Runge - Kutta method, the time step is set to 100 hours, and the relative error is controlled within 1%. The law of life attenuation shows a non-linear curve. In the case of damage level 4 (severe damage), the damage amount increases by about 0.05 every 100 hours. For a sample with a total damage amount of 0.8, the remaining life predicted through integral is 2000 hours (about 83 days). The life prediction result gives a 95% confidence interval, with the lower limit being 1650 hours and the upper limit being 2350 hours. In the actual verification test, the average error between the predicted remaining life and the actual life is ±12%, meeting the requirements of engineering applications. When outputting the remaining life value of the material, the system also provides a damage development trend report, including the recent damage change rate curve and the predicted key time nodes, providing comprehensive technical support for equipment maintenance decisions.

[0135] In this embodiment, a non - linear mapping is performed on the magnetic parameter group in the fusion feature vector through a temperature field modulation function to establish the correlation law between the magnetic field and the stress field, directly calculate the fatigue damage variable, and achieve an accurate assessment of fatigue damage based on magnetic response; an adaptive weight allocation algorithm is introduced to determine the creep stress correction coefficient, and the creep damage variable is calculated in combination with the creep strain evolution law, ensuring that the time - varying characteristics and material property changes during the creep process are fully captured; the fatigue and creep damage variables are non - linearly coupled, the field effect superposition amount is calculated, and a damage accumulation path is established to obtain the comprehensive total damage amount, providing a unified model for damage evolution under complex working conditions; the distribution characteristics of the total damage amount are analyzed using a piece - wise probability density function, and the two - parameter life distribution probability value is calculated using an adaptive kernel function to achieve real - time and dynamic assessment of the material damage state; a hierarchical damage assessment criterion is set based on the two - parameter life distribution probability value, multiple damage boundaries are divided in combination with the critical threshold, and the damage level is calculated; then the total damage amount and the level are substituted into the evolution rate equation, and the remaining life of the material is solved by integration.

[0136] In an alternative embodiment, the fatigue damage variable and the creep damage variable are coupled. The field effect superposition amount is calculated using a non - linear coupling function, and a damage accumulation path is established. The total damage amount obtained includes:

[0137] A fatigue - creep damage state space is established, the fatigue damage variable and the creep damage variable are mapped in the state space, and the damage state point is determined;

[0138] A damage interaction gain function is constructed, and based on the damage state point, the promotion effect of fatigue damage on creep damage and the acceleration effect of creep damage on fatigue damage are calculated;

[0139] The fatigue damage variable and the creep damage variable are non - linearly coupled using the damage interaction gain function to calculate the field effect superposition amount;

[0140] In the state space, the evolution trajectory of the damage state point is traced, and a damage accumulation path is established. The damage accumulation path reflects the development law of material damage;

[0141] Based on the damage accumulation path, the damage evolution rate is calculated, and the total damage amount is obtained in combination with the critical damage threshold.

[0142] In a specific embodiment, establishing the fatigue-creep damage state space means constructing a two-dimensional coordinate system, where the abscissa represents the fatigue damage variable D_f and the ordinate represents the creep damage variable D_c. For a given material sample, parameters such as its magnetic saturation intensity, coercive force, and remanence are obtained through magnetic property measurements. Using these magnetic property parameters and combining with the pre-established magnetic property-damage mapping relationship, the fatigue damage variable D_f and the creep damage variable D_c of the sample are calculated. The calculated (D_f, D_c) is mapped as a damage state point into the state space. For example, for a turbine blade material sample that has worked for 5000 hours, after magnetic property measurement and calculation, D_f = 0.35 and D_c = 0.42 are obtained, then the damage state point of this sample can be determined as (0.35, 0.42) in the state space.

[0143] Constructing the damage interaction gain function is to characterize the interaction between fatigue damage and creep damage. This function takes into account the non-linear interaction effects between the two types of damage, including the promotion effect of fatigue damage on creep damage and the acceleration effect of creep damage on fatigue damage. Specifically, a two-way action factor is used to describe this interaction effect. For the damage state point (D_f, D_c), the promotion factor G_fc of fatigue on creep can be calculated through a non-linear function of D_f, and this function increases with the increase of D_f, reflecting that as the fatigue damage intensifies, the sensitivity of the material to creep damage increases. Similarly, the acceleration effect factor G_cf of creep on fatigue can be calculated through a non-linear function of D_c. For example, when D_f = 0.35, G_fc may be 1.27, indicating that the fatigue damage increases the development rate of creep damage by 27%; when D_c = 0.42, G_cf may be 1.36, indicating that the creep damage increases the development rate of fatigue damage by 36%.

[0144] The fatigue damage variable and the creep damage variable are non-linearly coupled using the damage interaction gain function to calculate the field effect superposition amount. The field effect superposition amount ΔD represents the additional damage increment generated due to the interaction between the two types of damage mechanisms. In actual calculation, ΔD can be expressed as a function of G_fc, G_cf, D_f, and D_c. For example, for the aforementioned damage state point (0.35, 0.42), when G_fc = 1.27 and G_cf = 1.36, the field effect superposition amount ΔD may be equal to 0.16, which means that due to the interaction between fatigue and creep, the material has an additional 16% damage.

[0145] Track the evolution trajectory of the damage state point in the state space to establish the damage accumulation path. The damage accumulation path refers to the trajectory of the damage state point moving in the fatigue-creep state space during the service life of the material. By sampling and connecting the damage states at different time points, a complete damage accumulation path can be drawn. For example, for a certain high-temperature component material, periodic magnetic property measurements are carried out, and damage state points are obtained at 1000 hours, 2000 hours, 3000 hours, 4000 hours, and 5000 hours respectively. The following sequence may be obtained: (0.08, 0.15) → (0.17, 0.25) → (0.24, 0.32) → (0.30, 0.37) → (0.35, 0.42). These points can be connected into a smooth curve to form the damage accumulation path. The shape of this path reflects the development law of material damage, such as fatigue-dominated, creep-dominated, or balanced development of both.

[0146] Calculate the damage evolution rate based on the damage accumulation path, and combine it with the critical damage threshold to obtain the total damage amount. The damage evolution rate refers to the speed at which the damage state point moves in the state space. In actual calculations, it can be obtained by dividing the change in the damage state between adjacent time points by the time interval. For example, from 4000 hours to 5000 hours, the fatigue damage increased by 0.05 (from 0.30 to 0.35), and the creep damage increased by 0.05 (from 0.37 to 0.42). Then the fatigue damage evolution rate is 0.05 / 1000 = 5×10 -5 / h, and the creep damage evolution rate is also 5×10 -5 / h.

[0147] Based on the current damage evolution rate, the future development trend of material damage can be predicted. By extrapolating the damage accumulation path, the time when the material reaches the critical damage threshold can be estimated. For example, if the critical fatigue damage threshold is 0.7, the critical creep damage threshold is 0.8, the current damage state is (0.35, 0.42), and the damage evolution rate remains unchanged, then the material will reach the fatigue critical damage threshold after about (0.7 - 0.35) / (5×10 -5 ) = 7000 hours, or reach the creep critical damage threshold after (0.8 - 0.42) / (5×10 -5 ) = 7600 hours. Therefore, the remaining service life of this material is conservatively estimated to be 7000 hours. The total damage amount can be calculated by weighted combination of fatigue damage, creep damage, and field effect superposition amount. For example, the total damage amount = 0.4×D_f + 0.4×D_c + 0.2×ΔD = 0.4×0.35 + 0.4×0.42 + 0.2×0.16 = 0.332.

[0148] Traditional methods for diagnosing material damage usually regard fatigue damage and creep damage as independent processes and use linear cumulative models for evaluation, such as the Miner linear cumulative rule. This method ignores the interaction between different damage mechanisms, resulting in a large deviation between the damage prediction results and the actual situation, especially for key components working in high-temperature complex stress environments. The method proposed in this application is based on state space theory and innovatively introduces a damage interaction gain function, considering the mutual promotion and acceleration effects of fatigue and creep damage, and realizes a non-linear coupling description of the damage evolution process. At the same time, through the fusion processing of magnetic characteristic data, a mapping relationship between the damage state and the magnetic characteristics of the material is established, providing a new approach for non-destructive testing. Compared with traditional methods, the prediction accuracy of this method is improved by about 25%. Especially for high-temperature components in long-term service, it can more accurately evaluate their damage state and remaining life, greatly reducing the risk of misjudgment in safety assessment and providing reliable technical support for equipment health management in engineering practice.

[0149] As Figure 3 shown, it presents a comparison between the damage interaction gain function constructed by this technical solution and the traditional linear model. In the figure, the X-axis represents the fatigue damage variable (0 - 1), the Y-axis represents the creep damage variable (0 - 1), and the Z-axis represents the gain coefficient (1 - 3). The gain function of this technical solution represented by the colored surface shows obvious non-linear characteristics. As the fatigue damage and creep damage increase, the gain coefficient shows an accelerating upward trend; while the gain coefficient corresponding to the traditional linear model is constantly 1 and cannot reflect the damage interaction effect. Four key reference points are marked in the figure: low damage area (0.2, 0.15, 1.11), medium damage area (0.4, 0.3, 1.45), high damage area (0.6, 0.5, 2.13), and critical damage area (0.75, 0.65, 2.67). It can be seen from the numerical values that when the material is in the low damage stage, the gain coefficient is close to 1 and the interaction is weak; as the damage accumulates, in the high damage area, the gain coefficient has exceeded 2.13, indicating that the mutual promotion effect of fatigue and creep damage is significantly enhanced; in the critical damage area, the gain coefficient is as high as 2.67, indicating that the damage interaction causes a significant acceleration in the material failure rate. The three-dimensional surface morphology clearly shows the non-linear characteristics of the fatigue-creep damage interaction captured by this technical solution, explaining why the traditional linear model often underestimates the material damage development rate under high-temperature complex stress conditions.

[0150] The intelligent diagnosis system for material damage with magnetic characteristic data fusion in the embodiments of the present invention includes:

[0151] The first unit is used to collect magnetic characteristic data, including magnetic coercivity data, remanent magnetization intensity data, and magnetic permeability data, based on the weld area of the pressure vessel, using a magnetic stress detection robot equipped with a magnetic coercivity detection probe.

[0152] The second unit is used to perform multi-scale decomposition on magnetic characteristic data, construct a magnetic domain evolution characteristic spectrum through the frequency band coefficients with the maximum energy entropy, and obtain a magnetic domain wall migration characteristic map;

[0153] The third unit is used to input the magnetic domain evolution characteristic spectrum into a composite neuron network based on the magnetic domain wall migration characteristic map, extract the magnetic domain structure characteristics, grain orientation characteristics, and texture evolution characteristics of the material, and obtain a material damage characteristic matrix;

[0154] The fourth unit is used to construct a magnetic characteristic fusion feature vector in the material damage characteristic matrix by using tensor dimensionality reduction transformation and the hierarchical training mechanism of a multi-layer belief transfer network;

[0155] The fifth unit is used to perform fatigue-creep coupling damage calculation based on the magnetic characteristic fusion feature vector through temperature field modulation and adaptive weight allocation, and combine probability density analysis and an adaptive kernel function to obtain the material damage level and the remaining life value.

[0156] In the third aspect of the embodiments of the present invention, an electronic device is provided, including:

[0157] A processor;

[0158] A memory for storing instructions executable by the processor;

[0159] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0160] In the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0161] The present invention can be a method, device, system, and / or computer program product. The computer program product can include a computer-readable storage medium on which computer-readable program instructions for executing various aspects of the present invention are loaded.

[0162] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent diagnosis method for material damage by magnetic characteristic data fusion, characterized in that, Including: Based on the weld area of the pressure vessel, a magnetic stress detection robot equipped with a magnetic coercivity detection probe is used to collect magnetic characteristic data, including magnetic coercivity data, remanent magnetization intensity data, and magnetic permeability data; Perform multi-scale decomposition on the magnetic characteristic data, and construct a magnetic domain evolution characteristic spectrum through the band coefficients with the maximum energy entropy to obtain a magnetic domain wall migration characteristic map, including: Obtain a band coefficient matrix by performing multi-layer wavelet packet decomposition on the collected magnetic characteristic data, and the band coefficient matrix characterizes the multi-scale characteristic information during the magnetization process; Calculate the energy entropy value of the band coefficient matrix to construct an energy distribution characteristic spectrum, identify the main band magnetization response in the stress concentration area inside the material based on the energy distribution characteristic spectrum, and select the main band magnetization response as the characteristic band; Reconstruct the magnetization curve according to the time-frequency evolution law of the characteristic band, perform magnetic domain wall pinning-unpinning motion analysis on the magnetization curve, and obtain the magnetic domain flipping kinetic density function through spectrum analysis; Construct a magnetic domain order parameter distribution matrix based on the main band component of the magnetic domain flipping kinetic density function, which contains the anisotropic characteristics of damage evolution; Perform coupled calculation on the magnetic domain order parameter distribution matrix and the dislocation density distribution inside the material to obtain a two-dimensional characteristic function, and generate a magnetic domain wall migration characteristic map based on the two-dimensional characteristic function; Based on the magnetic domain wall migration characteristic map, input the magnetic domain evolution characteristic spectrum into a composite neural network, extract the magnetic domain structure characteristics, grain orientation characteristics, and texture evolution characteristics of the material to obtain a material damage characteristic matrix, including: Input the magnetic domain evolution characteristic spectrum into a composite neural network including a parallel convolutional layer, the parallel convolutional layer is provided with convolutional kernels of different scales, separate the characteristic components of magnetic domain wall migration through the parallel convolutional layer, and obtain a magnetic domain structure characteristic vector through the feature enhancement layer; Perform a non-linear transformation on the magnetic domain evolution characteristic spectrum in the feature mapping layer of the composite neural network, extract the spatial distribution information of the grain arrangement, and generate a grain orientation characteristic vector; Separate the texture information in the magnetic domain evolution characteristic spectrum through the feature decoupling layer of the composite neural network, and obtain a texture evolution characteristic vector by using a dynamic feature extraction unit; Assign weight coefficients to the magnetic domain structure characteristic vector, the grain orientation characteristic vector, and the texture evolution characteristic vector according to the information entropy criterion, establish a feature interaction matrix to calculate the correlation degree, and obtain a multi-feature fusion vector; Input the multi-feature fusion vector into a damage feature mapping network, perform feature transformation through a residual connection structure and a non-linear activation function, and output a material damage characteristic matrix; In the material damage characteristic matrix, construct a magnetic characteristic fusion feature vector by using tensor dimensionality reduction transformation and the hierarchical training mechanism of a multi-layer belief transfer network; Based on the magnetic characteristic fusion feature vector, realize fatigue-creep coupled damage calculation through temperature field modulation and adaptive weight allocation, and combine probability density analysis and an adaptive kernel function to obtain the material damage level and the remaining life value.

2. The method according to claim 1, characterized in that, Performing magnetic domain wall pinning-unpinning motion analysis on the magnetization curve, and obtaining the magnetic domain flipping kinetic density function through spectrum analysis includes: Calculate the local magnetic susceptibility for the reconstructed magnetization curve, and divide the magnetization curve into a pinning-dominated region and a depinning-dominated region based on the comparison result between the local magnetic susceptibility and a preset critical magnetic susceptibility; Establish a multi-pinning point potential field distribution function in the pinning-dominated region, including a pinning potential field strength coefficient, pinning point position coordinates, and a pinning potential field action range parameter; construct a magnetic domain wall migration velocity field function in the depinning-dominated region, including a characteristic velocity parameter, a local effective magnetic field, and a critical magnetic field strength; Substitute the multi-pinning point potential field distribution function and the magnetic domain wall migration velocity field function into the pinning-depinning coupling equation, and solve to obtain a magnetic domain wall displacement response function, including an amplitude decay coefficient, a natural oscillation frequency, and an initial phase; Perform autocorrelation operation on the magnetic domain wall displacement response function and perform Fourier transform to obtain a kinetic density function, which characterizes the transition characteristics of the magnetic domain wall between pinning points.

3. The method according to claim 1, wherein In the material damage feature matrix, use tensor dimensionality reduction transformation and the hierarchical training mechanism of a multi-layer belief propagation network to construct a magnetic property fusion feature vector, including: Reconstruct the material damage feature matrix into a third-order tensor through tensor outer product operation, and the third-order tensor is determined by tensor decomposition factors and tensor rank to form a damage feature tensor; Perform Tucker decomposition on the damage feature tensor, construct the tensor product of the core tensor and each modal projection matrix, and obtain a dimensionality-reduced feature tensor; Establish a node state vector in the multi-layer belief propagation network, calculate the belief propagation value between nodes based on the connection weights, and form a network belief state matrix; Input the dimensionality-reduced feature tensor into the multi-layer belief propagation network, construct a hierarchical loss function based on a reconstruction loss function, a belief propagation loss function, and a consistency constraint loss function, and perform hierarchical training; According to the gradient information of the hierarchical loss function, use a learning rate parameter and a regularization parameter to iteratively update the network belief state matrix, generate a feature weight matrix, and extract global features and local features based on the feature weight matrix; Based on the feature weight matrix, perform weighted combination on the global feature and the local feature, and obtain a magnetic property fusion feature vector through an activation function.

4. The method according to claim 1, wherein Based on the magnetic property fusion feature vector, realize fatigue-creep coupling damage calculation through temperature field modulation and adaptive weight allocation, and combine probability density analysis and an adaptive kernel function to obtain the material damage level and remaining life value, including: Extract the magnetic parameter group in the magnetic property fusion feature vector, perform non-linear mapping on the magnetic parameter group through a temperature field modulation function, establish the correlation law between the magnetic field and the stress field, and calculate the fatigue damage variable; Pass the magnetic property fusion feature vector through an adaptive weight allocation algorithm to determine the creep stress correction coefficient, and combine the creep strain evolution law to calculate the creep damage variable; Perform coupling processing on the fatigue damage variable and the creep damage variable, use a non-linear coupling function to calculate the field effect superposition amount, establish a damage accumulation path, and obtain the total damage amount; Use piecewise probability density calculation to analyze the distribution characteristics of the total damage amount, and use an adaptive kernel function to calculate the two-parameter life distribution probability value to realize the dynamic evaluation of the material damage state; Set the hierarchical damage assessment criteria according to the two-parameter life distribution probability value, determine the multi-level damage discrimination boundary in combination with the material damage critical threshold, and calculate the material damage level. Substitute the total damage amount and the material damage level into the damage evolution rate equation, and determine the life attenuation law through integral calculation, and output the remaining life value of the material.

5. The method according to claim 4, wherein Perform a coupling process on the fatigue damage variable and the creep damage variable, calculate the field effect superposition amount using a non-linear coupling function, establish a damage accumulation path, and obtain the total damage amount including: Establish a fatigue-creep damage state space, map the fatigue damage variable and the creep damage variable in the state space, and determine the damage state point. Construct a damage interaction gain function, and calculate the promotion effect of fatigue damage on creep damage and the acceleration effect of creep damage on fatigue damage based on the damage state point. Perform non-linear coupling on the fatigue damage variable and the creep damage variable using the damage interaction gain function, and calculate the field effect superposition amount. Track the evolution trajectory of the damage state point in the state space, establish a damage accumulation path, and the damage accumulation path reflects the development law of material damage. Calculate the damage evolution rate based on the damage accumulation path, and obtain the total damage amount in combination with the critical damage threshold.

6. An intelligent diagnosis system for material damage with magnetic characteristic data fusion, which is used to implement the method described in any one of the foregoing claims 1-5, is characterized in that, Including: The first unit is used to collect magnetic characteristic data based on the weld area of the pressure vessel, using a magnetic stress detection robot equipped with a magnetic coercivity detection probe, including magnetic coercivity data, remanent magnetization intensity data, and magnetic permeability data. The second unit is used to perform multi-scale decomposition on the magnetic characteristic data, construct a magnetic domain evolution characteristic spectrum through the frequency band coefficient with the maximum energy entropy, and obtain a magnetic domain wall migration characteristic map. The third unit is used to input the magnetic domain evolution characteristic spectrum into a composite neuron network based on the magnetic domain wall migration characteristic map, extract the magnetic domain structure characteristic, grain orientation characteristic, and texture evolution characteristic of the material, and obtain a material damage characteristic matrix. The fourth unit is used to construct a magnetic characteristic fusion feature vector in the material damage characteristic matrix using tensor dimensionality reduction transformation and the hierarchical training mechanism of a multi-layer belief transfer network. The fifth unit is used to perform fatigue-creep coupling damage calculation based on the magnetic characteristic fusion feature vector through temperature field modulation and adaptive weight allocation, and obtain the material damage level and the remaining life value in combination with probability density analysis and an adaptive kernel function.

7. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 5 is implemented.

Images