Detection method for defects in three-layer conductive structures

Abstract

The present invention relates to a detection method for defects in a three-layer conductive structure. The method uses Fourier transform and Hilbert transform to analyze the frequency domain and instantaneous frequency domain of a pulsed eddy current detection signal and extract features, and extracts features such as peak value and zero-crossing time in the signal time domain, so as to obtain defect information of the three-layer conductive structure better and comprehensively; and adopts a feature dimension reduction method based on AIC and Fisher to reduce feature redundancy, so that the reduced-dimensional features can better characterize the defect type, improve the recognition accuracy and efficiency of the classifier, and finally use SVM to construct a classifier, and use a particle swarm algorithm to optimize the classifier model parameters, so as to improve the training speed and accuracy. The present invention can improve the precision and recall rate of defect position and defect size recognition of the three-layer conductive structure, and can reduce the error of defect size recognition.

Description

Technical field:

[0001] The invention relates to a defect nondestructive detection technology, and in particular to a detection method for defects in a three-layer conductive structure. Background technology:

[0002] Multilayer conductive structures are widely used in the manufacture of aircraft, nuclear power plants, rails, etc., and are important industrial manufacturing materials. During long-term use, multilayer conductive structures may be subjected to some external forces such as fatigue loads, impacts and extrusion, resulting in crack defects at the rivet joints, which may affect the safety of the use of multilayer conductive structures, causing economic losses and even threatening life. Therefore, it is necessary to evaluate the safety of the use of multilayer conductive structures, that is, it is necessary to detect the location of the cracks and quantitatively analyze the cracks, estimate the size of the cracks, and then carry out targeted maintenance. At present, the commonly used eddy current nondestructive testing technologies for multilayer conductive structures include pulsed eddy current, conventional eddy current, far-field eddy current and other testing technologies. Compared with other eddy current testing technologies such as conventional eddy current and far-field eddy current, pulsed eddy current testing technology has richer information in the frequency domain and can effectively reduce the influence of skin effect and lift-off effect. Therefore, pulsed eddy current testing technology is selected to detect multilayer conductive structures, obtain and analyze pulsed eddy current signals that can penetrate multilayer conductive structures, and thus estimate the location and size of defects.

[0003] At present, some scholars at home and abroad focus on the research of multi-layer conductive structure defect detection methods based on pulsed eddy current, so as to estimate the defect location: for example, Hosseini S et al. used Rihaczek distribution to analyze the frequency domain characteristics of the signal, combined with the time domain characteristics, and used principal component analysis to extract features to determine the distribution of defects in the multi-layer conductive structure; Chen X et al. used principal component analysis to select features, and then used Fisher linear discriminant method to distinguish the surface and sub-surface defects of the third layer; He Yunze et al. used principal component analysis to select features, and then used support vector machine to distinguish the surface and sub-surface defects of the first and second layers. However, these scholars only studied the surface or sub-surface defect locations of single-layer, one-layer or two-layer conductive structures, and rarely studied the situation where defects exist in each layer of the three-layer conductive structure.

[0004] Another group of scholars estimated the defect size of a certain layer in a multi-layer conductive structure through eddy current detection. For example, Abidin IZ used excitation signals with different pulse widths to detect multi-layer conductive structures and extracted the frequency components of the corresponding pulse eddy current detection signals as features, indicating that this feature can reflect the change of defect size to a certain extent. Although the features extracted in this document can reflect the size information, there is no specific quantitative analysis of the defect size; Zhang Siquan used the finite element method to quantitatively analyze the corrosion defects hidden in the multi-layer conductive structure; Stott CA used principal component analysis and clustering algorithms to quantitatively analyze the defects of the second layer of the multi-layer conductive structure. However, the above scholars only quantitatively analyzed the defect size of a certain layer.

[0005] In addition, none of the above scholars have comprehensively considered the problem of defect location and size. Therefore, another small number of scholars have comprehensively studied the problem of defect location and size estimation. For example, Huang P uses projection tracking based on information divergence index to extract signal features for defect location and quantification. However, when this document uses projection tracking based on information divergence index to reduce the dimension of features, it is easily affected by redundant and invalid features, resulting in poor results in defect location and quantitative analysis based on the new features obtained by dimensionality reduction. In addition, the principal component analysis method proposed by some scholars is an unsupervised feature extraction method that does not make good use of prior knowledge of features for identification.

[0006] In short, currently scholars rarely involve the analysis of defect location and size under the three-layer conductive structure, and only conduct quantitative analysis on the defect size existing in a certain layer. There are shortcomings such as poor effect of dimensionality reduction algorithm. In view of this, this case was born. Summary of the invention:

[0007] The present invention proposes a detection method for defects in a three-layer conductive structure. By extracting 47-dimensional feature parameters and adopting a feature dimensionality reduction method based on AIC and Fisher, the optimal feature dimensionality reduction dimension is found. Finally, a classifier is constructed using a support vector machine (SVM), and a particle swarm algorithm is used to optimize the classifier model parameters, thereby improving the recognition accuracy of defect position and size and reducing model training time.

[0008] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is:

[0009] The method for detecting defects in a three-layer conductive structure includes the following steps:

[0010] Step 1: Prepare a defect-free specimen and several defective specimens with a three-layer conductive structure. The defect in each defective specimen is set only in one layer. Use pulsed eddy current testing equipment to test the prepared specimens, obtain pulsed eddy current testing signal data, and use it as a training sample to construct a defect position recognition training set and a defect size recognition training set;

[0011] Step 2: Parameter initialization setting: Initialize the independent parameter coefficient λ and the current mode, and set the dimension reduction threshold K yu and defect training model flag FLQ, let FLQ = 0, current dimension K = 1;

[0012] Step 3: Read the data in the corresponding mode according to the current mode, and convert the pulse eddy current signal into a time domain signal to extract the time domain characteristic parameters; perform Fourier transform on the time domain signal to extract the frequency domain characteristic parameters; decompose the time domain signal and perform Hilbert transform to extract the instantaneous frequency domain characteristic parameters; combine the extracted time domain, frequency domain and instantaneous frequency domain characteristic parameters of the pulse eddy current signal to construct the feature set Λ;

[0013] Step 4: Determine the current mode: If the current mode is "defect location training mode" or "defect size training mode", continue to step 5 and subsequent steps; if the current mode is "detection mode", jump directly to step 8;

[0014] Step 5: Use Fisher's method to reduce the feature set Λ to K dimensions, calculate and record the AIC value corresponding to the K dimension;

[0015] Step 6: Determine whether K is greater than the threshold K yu :If K is greater than the threshold K yu , then take the Fisher projection dimension K corresponding to the minimum AIC value best , extract K best Feature set Λ' and record projection direction under dimensional projection Then, according to the current mode, SVM is used to construct the corresponding classifier; if K is less than or equal to the threshold K yu , then K=K+1, and return to step 5 to perform dimensionality reduction processing;

[0016] Step 7: Train the classifier, use the particle swarm method to optimize the penalty factor and Gaussian function kernel parameters, complete the training, and save the trained classifier;

[0017] Determine the current mode: If the current mode is "defect location training mode", change the current mode to "defect size training mode", set FLQ = 1, K = 1, and then jump to step 3; if the current mode is "defect size training mode" and FLQ < 4, set FLQ = FLQ + 1, K = 1, and jump to step 3; if the current mode is "defect size training mode" and FLQ = 4, change the current mode to "detection mode", and set the projection matrix Skip to step 3;

[0018] Step 8: Feature Set Λ Perform projection and read the projected feature set Λ ", call the defect position recognition classifier to identify the defect position and output it. If the recognition result is no defect, jump to step 3; if the recognition result is defective, re-project the feature set Λ, call the defect size recognition classifier to identify the defect size, output the defect size and re-execute step 3.

[0019] Furthermore, there are a total of 19 specimens in step 1, including 1 defect-free specimen and 18 defective specimens, and the defect sizes are set to the following 6 types: 2mm, 3mm, 4mm, 5mm, 6mm, and 7mm. The defects of the 6 defect sizes are respectively set in the first layer of conductive structure to form 6 first-layer defective specimens; the defects of the 6 defect sizes are respectively set in the second layer of conductive structure to form 6 second-layer defective specimens; the defects of the 6 defect sizes are respectively set in the third layer of conductive structure to form 6 third-layer defective specimens.

[0020] Furthermore, among the pulsed eddy current detection signal data obtained, 200 groups of pulsed eddy current detection signal data of each defective specimen and 350 groups of pulsed eddy current detection signal data of defect-free specimens are obtained, and each group has 2000 sampling data; 350 groups of defect-free signal data with 2000 items in each group, and 50 groups of defective specimen signal data with 2mm, 3mm, 4mm, 5mm, 6mm and 7mm are randomly selected from each layer, and 2000 items in each group are used as training sets for defect position identification; 200×2000 signal data are selected for each defect size of 2mm, 3mm, 4mm, 5mm, 6mm and 7mm in each of the three layers, so 6×200 groups of signal data with 2000 items in each group are obtained for each layer as training sets for defect size identification.

[0021] Further, if the current mode in step 3 is "defect position training mode", the training set for defect position identification is read; if the current mode is "defect size training mode" and when FLQ=1, the training set for defect size identification of the first layer is read; when FLQ=2, the training set for defect size identification of the second layer is read; when FLQ=3, the training set for defect size identification of the third layer is read; if the current mode is "detection mode", the pulse eddy current signal data of the three-layer conductive structure directly detected is read.

[0022] Furthermore, the time domain characteristic parameters extracted in step 3 are respectively the peak value, zero crossing time and peak time of the pulse eddy current signal; the extracted frequency domain characteristic parameters are the fundamental frequency component and the second to twelfth harmonic components of the pulse eddy current signal; the time domain signal of the pulse eddy current signal is decomposed by the empirical mode decomposition algorithm to obtain four intrinsic mode functions, and the four intrinsic mode functions are Hilbert transformed. The extracted time and frequency domain characteristic parameters include the mean, instantaneous phase variance, instantaneous frequency variance, instantaneous amplitude variance of the four intrinsic mode functions of the pulse eddy current signal, and the bandwidth, variance, area, and maximum value of the marginal spectrum.

[0023] Further, the Fourier transform is performed according to formula (1):

[0024]

[0025] Among them, F(w h ) indicates w h The frequency domain signal, w h Indicates the frequency value, represents an imaginary number, f(t) represents the amplitude function of the pulsed eddy current signal, and t represents time;

[0026] The Hilbert transform is performed according to formula (2):

[0027]

[0028] Among them, c d (t) represents the dth eigenmode function, represents the function of the dth intrinsic mode function after Hilbert transformation, and * represents convolution;

[0029] The instantaneous amplitudes of the four eigenmode functions are calculated by formula (3), the instantaneous phases of the four eigenmode functions are calculated by formula (4), the instantaneous frequencies of the four eigenmode functions are calculated by formula (5), and the marginal spectra of the four eigenmode functions are calculated by formula (6):

[0030]

[0031]

[0032] w d (t) = dθ d (t) / dt (5)

[0033]

[0034] Among them, a d (t) represents the instantaneous amplitude of the dth eigenmode function, θ d (t) represents the instantaneous phase of the dth eigenmode function, w d (t) represents the instantaneous frequency of the dth eigenmode function, H(w d ) represents the marginal spectrum of the dth intrinsic mode function, and Re() represents a real number.

[0035] Furthermore, the dimension reduction method in step 5 is as follows:

[0036] The intra-class scatter matrix S of the feature samples is calculated by formula (7): W , calculate the inter-class scatter matrix S by formula (8) B :

[0037]

[0038]

[0039] Where C represents the number of defect categories, x ji represents the i-th sample of the j-th category, m j represents the number of samples in the jth category, u j represents the sample mean of the jth category, u represents the mean of all samples, and p j represents the prior experience of the jth category;

[0040] By Fisher's dimensionality reduction criterion J(W K ) reduces the feature samples to K dimensions, J(W K ) is expressed as:

[0041]

[0042] Among them, W K ={w1,w2,...,w K} represents the projection direction of each feature;

[0043] The AIC value is calculated as follows:

[0044] The intra-class deviation of each K-dimensional sample is calculated by formula (10):

[0045]

[0046] Among them, D i represents the intra-class deviation of sample i, represents the k-th eigenvalue of sample i after dimensionality reduction, represents the k-th eigenvalue of the center of the class to which sample i belongs after dimensionality reduction; considering that the intra-class deviation of sample features follows Gaussian distribution, the maximum likelihood estimation function of the intra-class deviation is calculated by formula (11):

[0047]

[0048] Among them, Γ represents the maximum likelihood estimation function, n represents the total number of samples, and u * represents the expected value, σ 2 represents the standard deviation;

[0049] Considering that the feature dimension will affect the time complexity of the SVM algorithm, the feature dimension reduction dimension K is taken as the independent parameter that needs to be considered in AIC. According to the maximum likelihood estimation function obtained from formula (11), formula (12) is constructed to calculate the AIC value corresponding to the K dimension:

[0050]

[0051] Among them, λ represents the independent parameter coefficient.

[0052] Furthermore, the classifier in step 6 is constructed according to the model in formula (13), as follows:

[0053]

[0054] Among them, Q(α) represents the classification model, α i ,α l denote the i-th and l-th Lagrange multipliers, y i ,y l Represent the i-th and l-th sample labels respectively, y i ,y l ∈{-1,1},M j represents the number of features of type j defect, Φ(x i ,x l ) represents the Gaussian function, and c represents the penalty factor.

[0055] Furthermore, if the current mode is the "defect location training mode", a defect location recognition classifier is constructed, and four classification models are constructed using model (13), namely SVM-0L, SVM-1L, SVM-2L, and SVM-3L, wherein SVM-0L uses the defect-free features in the training feature set Λ' as the positive sample set, and the other features as the negative sample set; SVM-1L uses the features of the first layer in the training feature set Λ' as the positive sample set, and the other features as the negative sample set; SVM-2L uses the features of the second layer in the training feature set Λ' as the positive sample set, and the other features as the negative sample set; SVM-3L uses the features of the third layer in the training feature set Λ' as the positive sample set, and the other features as the negative sample set, and saves

[0056] If the current mode is "defect size training mode" and FLQ=1, save Construct the first-level defect size recognition classifier, and use model (13) to construct the same number of classification models as the defect size types set in the specimen; if the current mode is "defect size training mode" and when FLQ = 2, save Construct the second-level defect size recognition classifier, and use model (13) to construct the same number of classification models as the defect size types set in the specimen; if the current mode is "defect size training mode" and when FLQ = 3, save Construct a third-level defect size recognition classifier, and use model (13) to construct classification models with the same number of defect size types as those set in the test piece.

[0057] Furthermore, the projection of the feature set Λ in step 8 is performed according to formula (14):

[0058] Λ"=W' K ×Λ (14)

[0059] When the defect location recognition classifier identifies the defect as the first layer, the parameter is read According to formula (14), the feature set Λ is re-projected to obtain the feature set Λ', and the first-layer defect size recognition classifier is called to perform defect size recognition; when the defect position recognition classifier recognizes the defect as the second-layer defect, the parameter is read. According to formula (14), the feature set Λ is re-projected to obtain the feature set Λ', and the second-layer defect size recognition classifier is called to perform defect size recognition; when the defect position recognition classifier recognizes the third-layer defect, the parameter is read. The feature set Λ is re-projected according to formula (14) to obtain the feature set Λ'' again, and the third-layer defect size recognition classifier is called to perform defect size recognition.

[0060] The beneficial effects of the present invention are mainly manifested in: the present invention proposes to use Fourier transform and Hilbert transform to analyze the frequency domain and instantaneous frequency domain of pulse eddy current detection signals and extract features, and extract features such as peak value and zero crossing time in the signal time domain, which can better and comprehensively obtain the defect information of the three-layer conductive structure. A feature dimension reduction method based on AIC and Fisher is proposed to reduce feature redundancy, so that the reduced features can better characterize the defect type and improve the recognition accuracy and efficiency of the classifier. Finally, the classifier is constructed using SVM, and the particle swarm algorithm is used to optimize the classifier model parameters, which can improve the training speed and accuracy. Finally, the present invention can improve the precision and recall rate of defect position and size identification, and reduce the error of defect size identification.

[0061] The present invention is further described below through the accompanying drawings and specific embodiments. Description of the drawings:

[0062] Figure 1 The figure is a flow chart of the defect detection method of the present invention. Specific implementation method:

[0063] This embodiment discloses a method for detecting defects in a three-layer conductive structure. Figure 1 As shown, the implementation process of this method is described as follows:

[0064] Step 1: Make a test piece and collect the pulsed eddy current detection signal data of the test piece as a training sample.

[0065] Among them, there are 19 specimens made in this embodiment, one of which is a defect-free specimen (that is, there are no defects in the three-layer conductive structure). The remaining 18 specimens are defective specimens, and the defect sizes are set as follows: 2mm, 3mm, 4mm, 5mm, 6mm, and 7mm. The 6 defect sizes given above are respectively set in the first layer of the 6 defective specimens, and it is ensured that the other layers of these 6 specimens are free of defects, forming specimens with 6 defect sizes whose defects are located in the first layer; according to the above method, the 6 defect sizes are respectively set in the second layer of the other 6 defective specimens, and it is ensured that the other layers are free of defects, at this time, specimens with 6 defect sizes located in the second layer are formed; the 6 defect sizes are respectively set in the third layer of the remaining 6 defective specimens, and it is ensured that the other layers are free of defects, at this time, specimens with 6 defect sizes located in the third layer are formed. This results in a defect-free specimen, 6 defect-size specimens with 6 defects located on the first layer, 6 defect-size specimens with 6 defects located on the second layer, and 6 defect-size specimens with 6 defects located on the third layer, for a total of 19 specimens.

[0066] The 19 specimens were tested by pulsed eddy current testing equipment, and 200 sets of pulsed eddy current test signal data for each defective specimen and 350 sets of pulsed eddy current test signal data for defect-free specimens were obtained, and each set had 2000 sampling data. 350 sets of defect-free signal data with 2000 data in each set and 50 sets of defective specimen signal data with 2mm, 3mm, 4mm, 5mm, 6mm and 7mm were randomly selected from each layer, and 2000 data in each group were used as training sets for defect position identification. 200×2000 signal data were selected for each defect size of 2mm, 3mm, 4mm, 5mm, 6mm and 7mm in each of the three layers, so 6×200 sets of signal data with 2000 data in each group were obtained for each layer as training sets for defect size identification.

[0067] Step 2: Parameter initialization setting: Initialize the independent parameter coefficient λ and the current mode, and set the dimension reduction threshold K yu is 47, the defect training model flag FLQ=0, and the current dimension K=1.

[0068] Step 3: Determine the current mode: the mode types are divided into "defect position training mode", "defect size training mode" and "detection mode"; if the current mode is "defect position training mode", read the training set for defect position identification, and then execute step 4; if the current mode is "defect size training mode", and when FLQ=1, read the training set for defect size identification of the first layer; when FLQ=2, read the training set for defect size identification of the second layer; when FLQ=3, read the training set for defect size identification of the third layer, and then execute step 4; if the current mode is "detection mode", read the pulse eddy current signal data of the three-layer conductive structure directly detected, and then execute step 4.

[0069] Step 4: Convert each group of pulsed eddy current signals into time domain signals and extract time domain characteristic parameters.

[0070] Before introducing the content of this step 4 in detail, let's first clarify a few concepts: "peak value" refers to the maximum voltage amplitude of the transient response signal detected simultaneously at the rising and falling edges of the pulsed eddy current signal; "zero crossing time" refers to the time interval from the rising or falling edge of the pulsed eddy current signal to the time when the voltage amplitude of the transient response signal drops to zero; "peak time" refers to the time interval from the rising or falling edge of the pulsed eddy current signal to the voltage peak. Convert each group of pulsed eddy current signals into time domain signals, and let the amplitude function of the pulsed eddy current signal be f(t), where t represents time, the peak value is max(f(t)), and let the time point when the peak value is reached be t fmax , find t fmax The last time point when f(t) = 0 is t1, and the time point when t fmaxThe first time point at which f(t) = 0 is t2, then the zero crossing time = t2-t1, and the peak time = t fmax -t1. Extract three characteristic parameters of each group of pulsed eddy current signals, namely, peak value, zero crossing time and peak time.

[0071] Step 5: Perform Fourier transform on the time domain signal in step 4 and extract frequency domain feature parameters.

[0072] Wherein, the time domain signal of each group of pulsed eddy current signals is subjected to Fourier transform as shown in formula (1).

[0073]

[0074] Among them, F(w h ) indicates w h The frequency domain signal, w h Indicates the frequency value, Represents an imaginary number. The fundamental frequency component F(2.5kHz) and the second to twelfth harmonic components of each group of pulse eddy current signals are extracted, namely F(5kHz), F(7.5kHz), F(10kHz), F(12.5kHz), F(15kHz), F(17.5kHz), F(20kHz), F(22.5kHz), F(25kHz), F(27.5kHz), and F(30kHz), with a total of 12 frequency domain feature parameters.

[0075] Step 6: Decompose the time domain signal and perform Hilbert transform.

[0076] The empirical mode decomposition algorithm is used to decompose the time domain signal of each group of pulsed eddy current signals to obtain four intrinsic mode functions. The four intrinsic mode functions are Hilbert transformed using formula (2).

[0077]

[0078] Among them, c d (t) represents the dth eigenmode function, represents the function of the dth intrinsic mode function after Hilbert transform, and * represents convolution.

[0079] Step 7: Extract instantaneous frequency domain feature parameters and construct feature set Λ.

[0080] The instantaneous amplitudes of the four intrinsic mode functions are calculated by formula (3), the instantaneous phases of the four intrinsic mode functions are calculated by formula (4), the instantaneous frequencies of the four intrinsic mode functions are calculated by formula (5), and the marginal spectra of the four intrinsic mode functions are calculated by formula (6).

[0081]

[0082]

[0083] w d (t) = dθ d (t) / dt (5)

[0084]

[0085] Among them, a d (t) represents the instantaneous amplitude of the dth eigenmode function, θ d (t) represents the instantaneous phase of the dth eigenmode function, w d (t) represents the instantaneous frequency of the dth eigenmode function, H(w d ) represents the marginal spectrum of the dth intrinsic mode function, and Re() represents a real number. The mean, instantaneous phase variance, instantaneous frequency variance, instantaneous amplitude variance and the bandwidth, variance, area and maximum value of the four intrinsic mode functions of each group of pulsed eddy current signals are extracted, totaling 32 instantaneous frequency domain feature parameters. The time domain, frequency domain and instantaneous frequency domain of each group of pulsed eddy current signals extracted are combined to form a feature set Λ.

[0086] Determine the current mode: If the current mode is "defect location training mode" or "defect size training mode", continue to execute step 8 and subsequent steps; if the current mode is "detection mode", jump directly to step 14.

[0087] Step 8: Use the Fisher method to reduce the feature set Λ to K dimensions. The specific dimensionality reduction method is as follows:

[0088] 8.1) Calculate the intra-class scatter matrix S of the feature samples using formula (7): W , calculate the inter-class scatter matrix S by formula (8) B :

[0089]

[0090]

[0091] Where C represents the number of defect categories, x ji represents the i-th sample of the j-th category, m j represents the number of samples in the jth category, u j represents the sample mean of the jth category, u represents the mean of all samples, and p j represents the prior experience of the jth category;

[0092] 8.2) Through Fisher's dimensionality reduction criterion J(W K ) can reduce the feature samples to K dimensions, which can be expressed as

[0093]

[0094] Among them, W K ={w1,w2,...,w K} represents the projection direction of each feature.

[0095] Step 9: Calculate and record the AIC value corresponding to the K dimension. The specific calculation method is as follows:

[0096] 9.1) Calculate the intra-class deviation of each K-dimensional sample using formula (10).

[0097]

[0098] Among them, D i represents the intra-class deviation of sample i, represents the k-th eigenvalue of sample i after dimensionality reduction, It represents the k-th eigenvalue of the center of the class to which sample i belongs after dimensionality reduction. Considering that the intra-class deviation of sample features follows Gaussian distribution, the maximum likelihood estimation function of the intra-class deviation is calculated by formula (11).

[0099]

[0100] Among them, Γ represents the maximum likelihood estimation function, n represents the total number of samples, and u * represents the expected value, σ 2 represents the standard deviation;

[0101] 9.2) Considering that the feature dimension will affect the time complexity of the SVM algorithm, the feature dimension reduction dimension K is taken as the independent parameter that needs to be considered in AIC. According to the maximum likelihood estimation function obtained from formula (11), formula (12) is constructed to calculate the AIC value corresponding to the K-dimensional dimension:

[0102]

[0103] Among them, λ represents the independent parameter coefficient.

[0104] Step 10: Determine whether K is greater than the threshold K yu :If K is greater than the threshold K yu , then take the Fisher projection dimension K corresponding to the minimum AIC value best , extract K best Feature set Λ' and record projection direction under dimensional projection Then execute step 11; if K is less than or equal to the threshold K yu , then K=K+1, and return to step 8.

[0105] Step 11: Determine the current mode and construct a classifier under the corresponding mode.

[0106] If the current mode is "defect location training mode", four classification models are constructed using model (13), namely SVM-0L, SVM-1L, SVM-2L, and SVM-3L as defect location recognition classifiers. Among them, SVM-0L uses the defect-free features in the training feature set Λ' as the positive sample set, and other features as the negative sample set; SVM-1L uses the features of the first layer in the training feature set Λ' as the positive sample set, and other features as the negative sample set; SVM-2L uses the features of the second layer in the training feature set Λ' as the positive sample set, and other features as the negative sample set; SVM-3L uses the features of the third layer in the training feature set Λ' as the positive sample set, and other features as the negative sample set, and saves

[0107] If the current mode is "defect size training mode" and FLQ=1, save Model (13) is used to construct six classification models, namely 1_SVM-2S, 1_SVM-3S, 1_SVM-4S, 1_SVM-5S, 1_SVM-6S, and 1_SVM-7S as the first-layer defect size recognition classifiers. Among them, 1_SVM-2S takes the features with the first layer defect size of 2mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 1_SVM-3S takes the features with the first layer defect size of 3mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 1_SVM-4S takes the features with the first layer defect size of 4mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 1_SVM-5S takes the features with the first layer defect size of 5mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 1_SVM-6S takes the features with the first layer defect size of 6mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 1_SVM-7S takes the features with the first layer defect size of 7mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set.

[0108] If the current mode is "defect size training mode" and FLQ=2, save Model (13) is used to construct six classification models, namely 2_SVM-2S, 2_SVM-3S, 2_SVM-4S, 2_SVM-5S, 2_SVM-6S, and 2_SVM-7S as the second-layer defect size recognition classifiers. Among them, 2_SVM-2S takes the features with a second-layer defect size of 2mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 2_SVM-3S takes the features with a second-layer defect size of 3mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 2_SVM-4S takes the features with a second-layer defect size of 4mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 2_SVM-5S takes the features with a second-layer defect size of 5mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 2_SVM-6S takes the features with a second-layer defect size of 6mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 2_SVM-7S takes the features with a second-layer defect size of 7mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set.

[0109] If the current mode is "defect size training mode" and FLQ=3, save W Kbest4 =W Kbest , model (13) is used to construct 6 classification models, namely 3_SVM-2S, 3_SVM-3S, 3_SVM-4S, 3_SVM-5S, 3_SVM-6S, and 3_SVM-7S as the third-layer defect size recognition classifier. Among them, 3_SVM-2S takes the features with the third layer defect size of 2mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 3_SVM-3S takes the features with the third layer defect size of 3mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 3_SVM-4S takes the features with the third layer defect size of 4mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 3_SVM-5S takes the features with the third layer defect size of 5mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 3_SVM-6S takes the features with the third layer defect size of 6mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set; 3_SVM-7S takes the features with the third layer defect size of 7mm in the training feature set Λ' as the positive sample set, and other features as the negative sample set.

[0110]

[0111] Among them, Q(α) represents the classification model, α i ,α l denote the i-th and l-th Lagrange multipliers, y i ,yl Represent the i-th and l-th sample labels respectively, y i ,y l ∈{-1,1},M j represents the number of features of type j defect, Φ(x i ,x l ) represents the Gaussian function, and c represents the penalty factor.

[0112] Step 12: Use the feature set Λ' to train the classifier, and use the particle swarm method to optimize parameters such as the penalty factor and Gaussian function kernel, complete the training, and save the classifier.

[0113] Step 13: Determine the current mode: If the current mode is "defect location training mode", change the current mode to "defect size training mode", set FLQ = 1, K = 1, and then jump to step 3; if the current mode is "defect size training mode" and FLQ < 4, set FLQ = FLQ + 1, K = 1, and jump to step 3; if the current mode is "defect size training mode" and FLQ = 4, change the current mode to "detection mode", set Skip to step 3.

[0114] Step 14: Using the projection matrix W' K , and project the feature set Λ according to formula (14);

[0115] Λ"=W' K ×Λ (14)

[0116] Step 15: According to the feature set Λ", the defect location recognition classifier is called to identify the defect location: if the recognition result is no defect, the result "currently no defect" is output and jump to step 3; if the recognition result is the first layer defect, the parameter is read Continue to step 16; if the identification result is a second layer defect, read the parameters Continue to step 16; if the identification result is a third layer defect, read the parameters Continue to step 16.

[0117] Step 16: Reproject the feature set Λ according to formula (14) to obtain the feature set Λ'' again, and call the corresponding defect size classifier to identify the defect size according to the defect position recognition result (if the defect position recognition result is a first-layer defect, call the first-layer defect size recognition classifier; if the defect position recognition result is a second-layer defect, call the second-layer defect size recognition classifier; if the defect position recognition result is a third-layer defect, call the third-layer defect size recognition classifier), then output the defect position and size of the three-layer conductive structure directly detected, and jump to step 3.

[0118] The above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Other modifications or equivalent substitutions made to the technical solution of the present invention by ordinary technicians in this field should be included in the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for detecting defects in a three-layer conductive structure, characterized in that: The steps include: Step 1: Prepare a defect-free specimen and several defective specimens with a three-layer conductive structure. The defect in each defective specimen is set only in one layer. Use pulsed eddy current testing equipment to test the prepared specimens, obtain pulsed eddy current testing signal data, and use it as a training sample to construct a defect position recognition training set and a defect size recognition training set; Step 2: Parameter initialization setting: Initialize the independent parameter coefficient λ and the current mode, and set the dimension reduction threshold K yu and defect training model flag FLQ, let FLQ = 0, current dimension K = 1; Step 3: Read the data in the corresponding mode according to the current mode, and convert the pulsed eddy current signal into a time domain signal to extract the time domain characteristic parameters; perform Fourier transform on the time domain signal to extract the frequency domain characteristic parameters; Decomposing the time domain signal and performing Hilbert transform to extract instantaneous frequency domain characteristic parameters; combining the extracted time domain, frequency domain and instantaneous frequency domain characteristic parameters of the pulse eddy current signal to construct a feature set Λ; Step 4: Determine the current mode: If the current mode is "defect location training mode" or "defect size training mode", continue to step 5 and subsequent steps; if the current mode is "detection mode", jump directly to step 8; Step 5: Use Fisher's method to reduce the feature set Λ to K dimensions, calculate and record the AIC value corresponding to the K dimension; Step 6: Determine whether K is greater than the threshold K yu :If K is greater than the threshold K yu , then take the Fisher projection dimension K corresponding to the minimum AIC value best , extract K best Feature set Λ' and record projection direction under dimensional projection Then, according to the current mode, SVM is used to construct the corresponding classifier; if K is less than or equal to the threshold K yu , then K=K+1, and return to step 5 to perform dimensionality reduction processing; Step 7: Train the classifier, use the particle swarm method to optimize the penalty factor and Gaussian function kernel parameters, complete the training, and save the trained classifier; Determine the current mode: If the current mode is "defect location training mode", change the current mode to "defect size training mode", set FLQ = 1, K = 1, and then jump to step 3; if the current mode is "defect size training mode" and FLQ < 4, set FLQ = FLQ + 1, K = 1, and jump to step 3; if the current mode is "defect size training mode" and FLQ = 4, change the current mode to "detection mode", set the projection matrix Skip to step 3; Step 8: Feature Set Λ Perform projection and read the projected feature set Λ ", call the defect position recognition classifier to identify the defect position and output it. If the recognition result is no defect, jump to step 3; if the recognition result is defective, re-project the feature set Λ, call the defect size recognition classifier to identify the defect size, output the defect size and re-execute step 3.

2. The method for detecting defects in a three-layer conductive structure according to claim 1, characterized in that: There are 19 specimens in the step 1, including 1 defect-free specimen and 18 defective specimens. The defect sizes are set to the following 6 types: 2mm, 3mm, 4mm, 5mm, 6mm, and 7mm. The defects of the 6 defect sizes are respectively set in the first layer conductive structure to form 6 first layer defective specimens; the defects of the 6 defect sizes are respectively set in the second layer conductive structure to form 6 second layer defective specimens; the defects of the 6 defect sizes are respectively set in the third layer conductive structure to form 6 third layer defective specimens.

3. The method for detecting defects in a three-layer conductive structure according to claim 2, characterized in that: Among the pulsed eddy current detection signal data obtained, 200 groups of pulsed eddy current detection signal data of each defective specimen and 350 groups of pulsed eddy current detection signal data of defect-free specimens are obtained, and each group has 2000 sampling data; 350 groups of defect-free signal data with 2000 items in each group, and 50 groups of defective specimen signal data with 2mm, 3mm, 4mm, 5mm, 6mm and 7mm are randomly selected from each layer, and 2000 items in each group are used as training sets for defect position identification; 200×2000 signal data are selected for each defect size of 2mm, 3mm, 4mm, 5mm, 6mm and 7mm in each of the three layers, so 6×200 groups of signal data with 2000 items in each group are obtained for each layer as training sets for defect size identification.

4. The method for detecting defects in a three-layer conductive structure according to claim 1, characterized in that: If the current mode in step 3 is "defect position training mode", the training set for defect position identification is read; if the current mode is "defect size training mode" and when FLQ=1, the training set for defect size identification of the first layer is read; when FLQ=2, the training set for defect size identification of the second layer is read; when FLQ=3, the training set for defect size identification of the third layer is read; if the current mode is "detection mode", the pulse eddy current signal data of the three-layer conductive structure directly detected is read.

5. The method for detecting defects in a three-layer conductive structure according to claim 1, characterized in that: The time domain characteristic parameters extracted in the step 3 are respectively the peak value, zero crossing time and peak time of the pulse eddy current signal; the frequency domain characteristic parameters extracted are the fundamental frequency component and the second to twelfth harmonic components of the pulse eddy current signal; the time domain signal of the pulse eddy current signal is decomposed by the empirical mode decomposition algorithm to obtain four intrinsic mode functions, and the four intrinsic mode functions are subjected to Hilbert transform. The extracted time and frequency domain characteristic parameters include the mean, instantaneous phase variance, instantaneous frequency variance, instantaneous amplitude variance of the four intrinsic mode functions of the pulse eddy current signal, and the bandwidth, variance, area and maximum value of the marginal spectrum.

6. The method for detecting defects in a three-layer conductive structure according to claim 5, characterized in that: The Fourier transform is performed according to formula (1): Among them, F(w h ) indicates w h The frequency domain signal, w h Indicates the frequency value, represents an imaginary number, f(t) represents the amplitude function of the pulsed eddy current signal, and t represents time; The Hilbert transform is performed according to formula (2): Among them, c d (t) represents the dth eigenmode function, represents the function of the dth intrinsic mode function after Hilbert transformation, and * represents convolution; The instantaneous amplitudes of the four eigenmode functions are calculated by formula (3), the instantaneous phases of the four eigenmode functions are calculated by formula (4), the instantaneous frequencies of the four eigenmode functions are calculated by formula (5), and the marginal spectra of the four eigenmode functions are calculated by formula (6): w d (t)=dθ d (t) / dt (5) Among them, a d (t) represents the instantaneous amplitude of the dth eigenmode function, θ d (t) represents the instantaneous phase of the dth eigenmode function, w d (t) represents the instantaneous frequency of the dth eigenmode function, H(w d ) represents the marginal spectrum of the dth intrinsic mode function, and Re() represents a real number.

7. The method for detecting defects in a three-layer conductive structure according to claim 1, characterized in that: The dimensionality reduction method in step 5 is as follows: The intra-class scatter matrix S of the feature samples is calculated by formula (7): W , calculate the inter-class scatter matrix S by formula (8) B : Where C represents the number of defect categories, x ji represents the i-th sample of the j-th category, m j represents the number of samples in the jth category, u j represents the sample mean of the jth category, u represents the mean of all samples, and p j represents the prior experience of the jth category; By Fisher's dimensionality reduction criterion J(W K ) reduces the feature samples to K dimensions, J(W K ) is expressed as: Among them, W K ={w1,w2,...,w K } represents the projection direction of each feature; The AIC value is calculated as follows: The intra-class deviation of each K-dimensional sample is calculated by formula (10): Among them, D i represents the intra-class deviation of sample i, represents the k-th eigenvalue of sample i after dimensionality reduction, represents the k-th eigenvalue of the center of the class to which sample i belongs after dimensionality reduction; considering that the intra-class deviation of sample features follows Gaussian distribution, the maximum likelihood estimation function of the intra-class deviation is calculated by formula (11): Among them, Γ represents the maximum likelihood estimation function, n represents the total number of samples, and u * represents the expected value, σ 2 represents the standard deviation; Considering that the feature dimension will affect the time complexity of the SVM algorithm, the feature dimension reduction dimension K is taken as the independent parameter that needs to be considered in AIC. According to the maximum likelihood estimation function obtained from formula (11), formula (12) is constructed to calculate the AIC value corresponding to the K dimension: Among them, λ represents the independent parameter coefficient.

8. The method for detecting defects in a three-layer conductive structure according to claim 1, characterized in that: The classifier in step 6 is constructed according to the model in formula (13), specifically as follows: Among them, Q(α) represents the classification model, α i ,α l denote the i-th and l-th Lagrange multipliers, y i ,y l Represent the i-th and l-th sample labels respectively, y i ,y l ∈{-1,1},M j represents the number of features of type j defect, Φ(x i ,x l ) represents the Gaussian function, and c represents the penalty factor.

9. The method for detecting defects in a three-layer conductive structure according to claim 8, characterized in that: If the current mode is "defect location training mode", a defect location recognition classifier is constructed, and four classification models are constructed using model (13), namely SVM-0L, SVM-1L, SVM-2L, and SVM-3L. Among them, SVM-0L uses the defect-free features in the training feature set Λ' as the positive sample set, and other features as the negative sample set; SVM-1L uses the features of the first layer in the training feature set Λ' as the positive sample set, and other features as the negative sample set; SVM-2L uses the features of the second layer in the training feature set Λ' as the positive sample set, and other features as the negative sample set; SVM-3L uses the features of the third layer in the training feature set Λ' as the positive sample set, and other features as the negative sample set. Save If the current mode is "defect size training mode" and FLQ=1, save Construct the first-level defect size recognition classifier, and use model (13) to construct the same number of classification models as the defect size types set in the specimen; if the current mode is "defect size training mode" and when FLQ = 2, save Construct the second-level defect size recognition classifier, and use model (13) to construct the same number of classification models as the defect size types set in the specimen; if the current mode is "defect size training mode" and when FLQ = 3, save Construct a third-level defect size recognition classifier, and use model (13) to construct classification models with the same number of defect size types as those set in the test piece.

10. The method for detecting defects in a three-layer conductive structure according to claim 1, characterized in that: The projection of the feature set Λ in step 8 is performed according to formula (14): Λ”=W' K ×L (14) When the defect location recognition classifier identifies the defect as the first layer, the parameter is read Reproject the feature set Λ according to formula (14) to obtain the feature set Λ'' again, and call the first layer defect size recognition classifier to perform defect size recognition; When the defect location recognition classifier identifies the defect as the second layer, the parameter is read Reproject the feature set Λ according to formula (14) to obtain the feature set Λ'' again, and call the second-layer defect size recognition classifier to perform defect size recognition; When the defect location recognition classifier identifies a third-layer defect, the parameter is read The feature set Λ is re-projected according to formula (14) to obtain the feature set Λ'' again, and the third-layer defect size recognition classifier is called to perform defect size recognition.

Images