A flip-chip vibration signal denoising method, system, device and medium

Abstract

The present application relates to a kind of flip chip vibration signal denoising method, system, equipment and medium, belong to flip chip vibration signal denoising technical field, wherein, method includes: the vibration signal of flip chip is collected;Initial sparse representation of the vibration signal is constructed, and sparse dictionary in the initial sparse representation is constructed according to the vibration signal, sparse representation model is constructed based on the sparse dictionary and the noise variance of vibration signal;The optimal sparse coefficient about the sparse dictionary in the sparse representation model is solved, and simultaneously the optimal sparse dictionary corresponding to the optimal sparse coefficient is obtained;Useless atom and noise atom in the optimal sparse dictionary are eliminated, and the optimal sparse dictionary after denoising;The vibration signal containing noise is reconstructed using the optimal sparse dictionary after denoising and the optimal sparse coefficient corresponding thereto, and the vibration signal after denoising is obtained.The present application can effectively denoise flip chip vibration signal.

Description

Technical Field

[0001] This invention relates to the field of flip chip vibration signal denoising technology, and in particular to a flip chip vibration signal denoising method, system, device and medium. Background Technology

[0002] Flip-chip bonding, as a high-performance electronic packaging technology, boasts advantages such as short interconnects, low connection resistance, and high reliability, making it an indispensable part of the electronics industry. However, as flip-chip technology continues to develop towards high-density solder joints and ultra-fine pitch, micro-defects such as cracks, voids, and missing balls often appear at the chip bonding points, severely affecting the reliability of chip packaging. Therefore, inspecting flip-chip bonding is of significant practical importance. Vibration testing is a commonly used non-destructive testing technique in defect detection and fault diagnosis. Its principle involves directly exciting the micro-interconnect solder joints of the flip-chip using an ultrasonic transducer. An air-coupled ultrasonic transducer is used to directly incident on the surface of the flip-chip at a 45° angle. Then, a Doppler laser vibrometer is used to detect the vibration on the chip surface. By distributively acquiring vibration displacement signals and performing signal analysis and processing, the reliability of the chip bonding can be determined. When facing advanced packaging defect detection, the complex chip packaging structure, the microscale of defects, the edge effect caused by complex defect boundaries, and the influence of material grain noise and detection system noise, etc., result in the useful signal being in a large-scale noise environment. Multiple modulation modes are mixed with each other, the signal components are complex, and the defect features are weak, which exacerbates the difficulty of packaging defect analysis and makes defect detection extremely challenging. Summary of the Invention

[0003] Therefore, the technical problem to be solved by the present invention is to overcome the problem of large-scale noise in the existing technology for detecting defects in flip chips.

[0004] To address the aforementioned technical problems, this invention provides a method for denoising vibration signals from flip-chip chips, comprising:

[0005] Step S1: Acquire vibration signals from the flip chip used for defect detection;

[0006] Step S2: Construct an initial sparse representation of the vibration signal, and construct a sparse dictionary in the initial sparse representation based on the vibration signal. Construct a sparse representation model based on the sparse dictionary and the noise variance of the vibration signal.

[0007] Step S3: Solve for the optimal sparse coefficients in the sparse representation model with respect to the sparse dictionary, and simultaneously obtain the optimal sparse dictionary corresponding to the optimal sparse coefficients;

[0008] Step S4: Remove useless and noise atoms from the optimal sparse dictionary to obtain the denoised optimal sparse dictionary;

[0009] Step S5: Reconstruct the noisy vibration signal using the denoised optimal sparse dictionary and its corresponding optimal sparse coefficients to obtain the denoised vibration signal.

[0010] In one embodiment of the present invention, the method for constructing the sparse representation model in step S2 includes:

[0011] The vibration signal is sparsely reconstructed to obtain an initial sparse representation:

[0012] ;

[0013] Where D is the sparse dictionary, S is the vibration signal matrix, N is the noise matrix, and X is the sparse coefficient of the vibration signal Y in the sparse dictionary D. This is an acceptable error term;

[0014] The initial sparse representation is transformed into an initial objective function, expressed as:

[0015] ;

[0016] in, To find the minimum value of the objective function, It is the l2 norm. The original signal matrix, The coefficient of the regularization term, It is a regularization term;

[0017] Constructing a sparse dictionary D using SVD decomposition: By applying SVD adaptive learning to the obtained vibration signal matrix S, it can be represented as:

[0018] ;

[0019] in, It is a left singular orthogonal matrix. It is a real matrix. for The number of feature vectors, for The i-th feature vector It is a right singular orthogonal matrix. for The number of feature vectors, for The i-th eigenvector, For singular value matrices with sparse properties, Singular value matrix The i-th singular value;

[0020] left singular orthogonal matrix As a sparse dictionary D, since the sparse dictionary D is adaptively learned from the vibration signal matrix S, it is used to effectively represent feature information;

[0021] Because the vibration signal from the flip chip is subject to noise interference during acquisition, the noise variance of the vibration signal matrix S is estimated using the median absolute deviation (MAD).

[0022] ;

[0023] in, For noise variance, This represents the median. To take the absolute value, c1 is the wavelet coefficient at the highest frequency scale layer of the wavelet decomposition;

[0024] Based on noise variance and singular value matrix Construct an adaptive dual weight matrix W, and set each element in the adaptive dual weight matrix W to:

[0025] ;

[0026] in, Singular value matrix The i-th singular value, i∈[1,2,...,M]; Let represent the noise variance of the j-th column of the vibration signal matrix, where j∈[1,2,....,N];

[0027] The adaptive dual weight matrix W is introduced into the initial objective function and used as a sparse representation model:

[0028] ;

[0029] in, For Hadama accumulation, It is an l1 norm.

[0030] In one embodiment of the present invention, the method for solving the optimal sparse coefficients of the sparse representation model with respect to the sparse dictionary in step S3 includes:

[0031] The sparse representation model is represented using the alternating direction multiplier method:

[0032] ;

[0033] in , , The constraint on X in the sparse representation;

[0034] Since the adaptive dual weight matrix W is a positive definite matrix that varies according to the vibration signal matrix S, and has no direct relationship with the sparse coefficients, it is extracted to obtain:

[0035] ;

[0036] in, This incorporates regularization terms that combine the dual weight matrix W;

[0037] To solve It can be expressed using the augmented Lagrange function, as shown in the formula:

[0038] ;

[0039] Where u represents the dual variable, ρ represents the transpose of the dual variable u; ρ represents the penalty parameter, which serves as the step size in the alternating direction multiplier method solution process.

[0040] To obtain the optimal Using the augmented Lagrange multiplier method (ALM) to refine the formula The iteration is as follows:

[0041] ;

[0042] ;

[0043] ;

[0044] The corresponding iterative solution is:

[0045] ;

[0046] ;

[0047] ;

[0048] in, X represents the sparse coefficients in the (k+1)th stage. Z represents the constraint variable Z of the sparse coefficients X in the k-th stage. Let u represent the dual variable in the k-th stage. Represents the solution in soft-threshold form. This represents the transpose of variable u in the k-th stage. Represents the identity matrix. Let X represent the sparse coefficients of the k-th stage. Abbreviated as And used as a soft threshold operator, the formula is:

[0049] ;

[0050] in, Indicates the first generation value. Indicates the second-generation value;

[0051] Based on the soft threshold operator The definition will The solution can be written in the following form:

[0052] ;

[0053] First scenario: When The solution satisfies At that time, bring it in From To study the convergence of the algorithm at this point, we will... or Eliminate; remove Iterative and Substitute it into this moment From:

[0054] ;

[0055] Assume Z* is The fixed point in the solution process, i.e. Then dual error At this point, we have:

[0056] ;

[0057] This represents the k-th stage of the dual error. express The dual error converges if and only if the spectral radius of matrix E is less than 1. To represent the eigenvalues ​​of matrix E, let... for The eigenvalues ​​of matrix E are then expressed as:

[0058] ;

[0059] This represents the i-th eigenvalue. Representing the eigenvalue representation of matrix E, when When the eigenvalue is greater than 0, the above equation is less than 1, and the algorithm guarantees convergence; therefore, solving for the optimal step size is equivalent to solving the following problem:

[0060] ;

[0061] because about It is monotonically decreasing, simplifying the above equation to:

[0062] ;

[0063] D represents T The smallest eigenvalue of D;

[0064] Second scenario: When The solution satisfies Then we get:

[0065] ;

[0066] Assume u* is The fixed point in the solution process, i.e. Then dual error At this point, we have:

[0067] ;

[0068] The algorithm converges at this point if and only if the spectral radius of matrix E is less than 1, and the eigenvalues ​​of matrix E are:

[0069] ;

[0070] when When the eigenvalue is greater than 0, the above expression is less than 1, and the algorithm guarantees convergence. about It is monotonically increasing, at this time:

[0071] ;

[0072] D represents T The largest eigenvalue of D;

[0073] Third scenario: When Z k+1 The solution satisfies Then:

[0074] ;

[0075] At this time, Z k+1 The matrix E is the same as in the first case, therefore we get:

[0076] ;

[0077] exist The iterative process of the three cases satisfies:

[0078] when When, the optimal step size ;

[0079] when When, the optimal step size ;

[0080] Will and This can be summarized in the following expression:

[0081] ;

[0082] and The corresponding spectral radius is expressed as:

[0083] ;

[0084] In the entire domain, the left half about Monotonically decreasing, right half about Monotonically increasing, therefore, assume If the optimal step size is located at the intersection of the two, then we get:

[0085] ;

[0086] Based on the obtained optimal step size Used to accelerate the augmented Lagrange multiplier method (ALM) for formulas Iteration is then performed to obtain the optimal sparsity coefficients. .

[0087] In one embodiment of the present invention, the method for removing useless atoms from the optimal sparse dictionary in step S4 includes:

[0088] The optimal sparse dictionary is used as the learning dictionary, which is an overcomplete dictionary. Useless atoms are removed from the overcomplete dictionary by using a dictionary mask, and then useful atoms are selected.

[0089] To reduce the number of atoms while keeping the errors in the reconstructed and vibrational signals below a preset value, the existing training data... , For the i-th sample, given an overcomplete dictionary have:

[0090] ;

[0091] in, To find the sparsity coefficients of the corresponding atoms in the overcomplete dictionary, To set the ith sparse coefficient corresponding to the overcomplete dictionary; A set of mask vectors, If the i-th atom is a useful atom, it needs to be retained. If the i-th atom is a useless atom and needs to be removed, then it is processed through the mask vector. The transformed sparsity coefficients are expressed as follows:

[0092] ;

[0093] in, For the mask vector The converted coefficients, For Hadama accumulation, For splicing operations, For the union, For the first to the Mth samples;

[0094] Simultaneously, the following conditions must be met:

[0095] .

[0096] In one embodiment of the present invention, the method of removing useless atoms from the optimal sparse dictionary in step S4 further includes: updating the atoms in the overcomplete dictionary after dictionary masking, sorting the absolute values ​​of the sparsity coefficients of each atom by gradient filtering, and removing atoms that have never been used or have been used less than a preset number of times.

[0097] In one embodiment of the present invention, the method for removing noise atoms from the optimal sparse dictionary in step S4 includes:

[0098] Treating each atom in the learning dictionary as an atomic component, the signal energy of the i-th atomic component in the learning dictionary is calculated using the following formula:

[0099] ;

[0100] The signal energy of each atomic component is normalized using the following formula:

[0101] ;

[0102] in, Let i be the signal energy of the i-th atomic component, i(t) be the t-th atomic component, and n be the number of all atomic components in the learning dictionary.

[0103] The formula for calculating the energy entropy increment of the signal energy of each atomic component is as follows:

[0104] ;

[0105] If the value of ΔQ(i) is higher than the preset discrimination threshold, it means that the atomic component is more authentic and the atomic component is retained; if the value of ΔQ(i) is lower than the preset discrimination threshold, it means that the atomic component is less authentic and the atomic component is treated as noise atom and removed.

[0106] To address the aforementioned technical problems, this invention provides a flip-chip vibration signal denoising system, comprising:

[0107] Acquisition module: Used to acquire vibration signals from flip chips used for defect detection;

[0108] Construction module: used to construct an initial sparse representation of the vibration signal, construct a sparse dictionary in the initial sparse representation based on the vibration signal, and construct a sparse representation model based on the sparse dictionary and the noise variance of the vibration signal.

[0109] Solving module: used to solve for the optimal sparse coefficients in the sparse representation model with respect to the sparse dictionary, and simultaneously obtain the optimal sparse dictionary corresponding to the optimal sparse coefficients;

[0110] Elimination module: used to remove useless and noisy atoms from the optimal sparse dictionary, resulting in a denoised optimal sparse dictionary;

[0111] Reconstruction module: Used to reconstruct the noisy vibration signal using the denoised optimal sparse dictionary and its corresponding optimal sparse coefficients, so as to obtain the denoised vibration signal.

[0112] To solve the above-mentioned technical problems, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the flip chip vibration signal denoising method described above.

[0113] To solve the above-mentioned technical problems, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the flip chip vibration signal denoising method described above are implemented.

[0114] To solve the above-mentioned technical problems, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-mentioned flip chip vibration signal denoising method.

[0115] The technical solution of the present invention has the following advantages compared with the prior art:

[0116] This invention presents a method for denoising vibration signals from flip-chips. Based on the vibration signal matrix S, a sparse dictionary D is constructed using SVD adaptive learning. An adaptive dual weight matrix W is constructed based on the noise variance of the vibration signal matrix S. Finally, a sparse representation model is obtained based on the sparse dictionary D and the adaptive dual weight matrix W. This invention removes useless atoms from the sparse dictionary D using a dictionary masking method and removes noise atoms by calculating the energy entropy increment corresponding to each atom, in order to find the best atom that can represent the vibration signal.

[0117] In solving for the optimal sparse coefficients of the sparse representation model, this invention sets the iteration step size in the augmented Lagrange function to speed up the iteration process of the augmented Lagrange multiplier method (ALM), obtaining the optimal sparse coefficients with the fewest iterations, effectively saving computation time and improving computational efficiency.

[0118] This invention can effectively denoise the vibration signal of flip chips and has good practicality. Attached Figure Description

[0119] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0120] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0121] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0122] Example 1

[0123] Reference Figure 1 As shown, this invention relates to a method for denoising vibration signals from a flip chip, comprising:

[0124] Step S1: Acquire vibration signals from the flip chip used for defect detection;

[0125] Step S2: Construct an initial sparse representation of the vibration signal, and construct a sparse dictionary in the initial sparse representation based on the vibration signal. Construct a sparse representation model based on the sparse dictionary and the noise variance of the vibration signal.

[0126] Step S3: Solve for the optimal sparse coefficients in the sparse representation model with respect to the sparse dictionary, and simultaneously obtain the optimal sparse dictionary corresponding to the optimal sparse coefficients;

[0127] Step S4: Remove useless and noise atoms from the optimal sparse dictionary to obtain the denoised optimal sparse dictionary;

[0128] Step S5: Reconstruct the noisy vibration signal using the denoised optimal sparse dictionary and its corresponding optimal sparse coefficients to obtain the denoised vibration signal.

[0129] It should be noted that the denoised optimal sparse dictionary and its corresponding optimal sparse coefficient (which can be considered as the second optimal sparse coefficient) in step S5 are different from the optimal sparse coefficient (which can be considered as the first optimal sparse coefficient) in step S3. The reason is that useless atoms and noise atoms have been removed from the optimal sparse dictionary in step S4. Since the number of atoms decreases, the optimal sparse coefficient (second optimal sparse coefficient) corresponding to the denoised optimal sparse dictionary in step S5 will automatically remove the correlation coefficients corresponding to useless atoms and noise atoms in the original optimal sparse coefficient (first optimal sparse coefficient).

[0130] The following is a detailed description of this embodiment:

[0131] Reference Figure 1 As shown, the present invention provides a method for denoising the vibration signal of a flip chip, which successfully separates the defect features of the chip vibration signal from the interference components / noise while ensuring the convergence efficiency of the dual sparse reconstruction algorithm.

[0132] Step S2, the steps for constructing the sparse representation model (i.e., an objective function) include:

[0133] S21: Sparsely reconstruct the vibration signal to obtain an initial sparse representation, which is expressed as:

[0134] ;

[0135] Where D is the sparse dictionary, S is the vibration signal matrix, N is the noise matrix, and X is the sparse coefficient of the vibration signal Y in the sparse dictionary D. This is an acceptable error term. A regularization term is introduced to transform the reconstructed sparse representation into the initial objective function:

[0136] ;

[0137] in, To find the minimum value of the objective function, It is the l2 norm. The original signal matrix, The coefficient of the regularization term, This is a regularization term.

[0138] Reconstruction errors based on the L2 norm are highly sensitive to noisy features. Therefore, if the data contains noise, the L2 norm term in minimizing the objective function will introduce noise into the representation. Both of these factors degrade the performance of subsequent feature selection, leading to redundant and noisy feature selection.

[0139] To address the problem of constructing a sparse dictionary for chip vibration signals under large-scale noise environments, we utilize SVD decomposition to construct a sparse dictionary D, i.e., applying SVD adaptive learning to the obtained vibration signal matrix S:

[0140] ;

[0141] in, It is a left singular orthogonal matrix. It is a real matrix. for The number of feature vectors, for The i-th eigenvector, It is a right singular orthogonal matrix. for The number of feature vectors, for The i-th eigenvector, For singular value matrices with sparse properties, Singular value matrix The i-th singular value. In this embodiment, the left singular orthogonal matrix is... As a sparse dictionary D, since it is adaptively learned from the vibration signal matrix S, it has strong flexibility and adaptive characteristics, and can effectively represent feature information.

[0142] Considering the time-varying characteristics of noise components and analyzing the distribution characteristics of noise interference that may exist during the chip vibration signal acquisition process, this embodiment uses the median absolute deviation (MAD) technique to estimate the noise variance of the vibration signal matrix S:

[0143] ;

[0144] in, This represents the median. To take the absolute value, c1 represents the wavelet coefficients at the highest frequency scale layer of the wavelet decomposition. Additionally, due to noise variance... The noise variance in the chip vibration signal fluctuates across different scales, so we consider using it as prior knowledge. We discover the intrinsic relationship between the dictionary space and the original data through coefficient reweighting, seeking stable representations of feature atoms in the dictionary. Based on the prior noise distribution in the chip vibration signal and the sparse distribution pattern of the singular value sequence of defect features, we construct an adaptive dual weighting matrix W for the latent representation of atoms. Each element in the dual weighting matrix W is set as follows:

[0145] ;

[0146] in, Singular value matrix The i-th singular value, i∈[1,2,...,M]; Let represent the noise variance of the j-th column of the signal matrix, where j∈[1,2,....,N].

[0147] The adaptive dual weight matrix W is introduced into the initial objective function and used as a sparse representation model:

[0148] ;

[0149] in, For Hadama accumulation, It is the l1 norm. The sparse representation model can effectively eliminate noise interference at different scales and retain the singular value component information corresponding to defect characteristics by weighting the sparse coefficient matrix X.

[0150] In step S3, the optimal sparsity coefficients for the sparse dictionary in the sparse representation model are solved, including:

[0151] Since ADMM (Alternating Direction Multiplier Method) possesses both the strong convergence property of the multiplier method and the decomposition property of the dual ascent method, it is an effective method for solving big data problems. The sparse representation model can be expressed using the alternating direction multiplier method as follows:

[0152] ;

[0153] in, , , This represents the constraint on X in the sparse representation.

[0154] Since the adaptive dual weight matrix W is a positive definite matrix that varies according to the vibration signal matrix S, and has no direct relationship with the sparse coefficients, it is proposed here:

[0155] ;

[0156] in, This is a regularization term that incorporates the dual weight matrix W.

[0157] Furthermore, in order to solve It is represented using the augmented Lagrangian function (here, for the purpose of solving). and However, it should be noted that the solution will be continuously solved during the iteration process, but the final iteration only needs to be solved. ):

[0158] ;

[0159] Where u represents the dual variable. ρ represents the transpose of the dual variable u, and ρ represents the penalty parameter (i.e., the step size in the ADMM solution process).

[0160] To obtain the optimal Using the augmented Lagrange multiplier method (ALM) to refine the formula The iteration is as follows:

[0161] ;

[0162] ;

[0163] ;

[0164] The corresponding iterative solution is:

[0165] ;

[0166] ;

[0167] ;

[0168] in, X represents the sparse coefficients in the (k+1)th stage. Z represents the constraint variable Z of the sparse coefficients X in the k-th stage. Let u represent the dual variable in the k-th stage. This represents the transpose of variable u in the k-th stage. Represents the identity matrix. X represents the sparse coefficients in the k-th stage. The solution is in the form of soft thresholding. Abbreviated as And it is defined as a soft threshold operator as follows:

[0169] ;

[0170] in, Indicates the first generation value (e.g.) ), Indicates the second generation value (e.g.) ).

[0171] Therefore, for ease of subsequent analysis, based on the definition of the soft threshold S above, Z is... k+1 The solution can be written in the following form (that is, the original solution). (Solution converted to three cases)

[0172] ;

[0173] The gradient method has a drawback: it is very sensitive to the choice of step size. Choosing an inappropriate step size has a significant impact on the number of iterations required for convergence or the convergence time, and may even cause the algorithm to fail to converge.

[0174] (1) Z k+1 First scenario:

[0175] When Z k+1 When the iteration is the first case, that is... Bring it into u k From To study the convergence of the algorithm at this point, let X... k or Z k Eliminate. Remove X. k+1 Iterative and Substitute Z at this time k+1 middle:

[0176] ;

[0177] Assume Z* is The fixed point in the solution process, i.e. Then dual error At this point, we have:

[0178] ;

[0179] Let the k-th stage of the dual error be represented. The dual error converges if and only if the spectral radius of matrix E is less than 1. To represent the eigenvalues ​​of E, let... for If the eigenvalues ​​are denoted by E, then the eigenvalues ​​of E can be expressed as:

[0180] ;

[0181] This represents the i-th eigenvalue. Representing the eigenvalue representation of matrix E, when When the eigenvalues ​​are greater than 0, the above expression is less than 1, and the algorithm guarantees convergence. Therefore, solving for the optimal step size is equivalent to solving the following problem:

[0182] ;

[0183] You can see about It is monotonically decreasing, so the above equation can be simplified to:

[0184] ;

[0185] D represents T The smallest eigenvalue of D.

[0186] (2) Z k+1 The second scenario:

[0187] When Z k+1 When the iteration satisfies the second condition, Z...k+1 =0, following the same steps, we can obtain:

[0188] ;

[0189] Assume u* is The fixed point in the solution process, i.e. Then dual error At this point, we have:

[0190] ;

[0191] The algorithm converges if and only if the spectral radius of matrix E is less than 1. The algorithm guarantees convergence, and the eigenvalues ​​of E are:

[0192] ;

[0193] Similarly, when When the eigenvalue is greater than 0, the above expression is less than 1. about It is monotonically increasing, at this time:

[0194] ;

[0195] D represents T The largest eigenvalue of D.

[0196] (3) Z k+1 The third scenario:

[0197] When Z k+1 When the iteration is the first case, that is... The analysis process is the same as the two methods above, so the main intermediate process results are listed directly here:

[0198] ;

[0199] It can be noted that Z at this time k+1 The matrix E is the same as in the first case, so the final result is given here:

[0200] .

[0201] The convergence states for all three cases have been given, consistent with what was found during the analysis process. and The optimal step size optimization problems in both cases are equivalent, so the optimal step size values ​​are the same. This is because they have the same convergence factor and the same matrix E; the difference lies in C, which does not affect the final convergence.

[0202] Therefore, in Zk+1 Through iterative processes involving the three scenarios, the following conclusions can be drawn:

[0203] when When, the optimal step size ;

[0204] when When, the optimal step size ;

[0205] Will and This can be summarized in the following expression:

[0206] ;

[0207] and The corresponding spectral radius can be expressed as:

[0208] ;

[0209] In the entire domain, the left half The right side of the equation is monotonically decreasing with respect to ρ, and monotonically increasing with respect to ρ. Therefore, assume that the optimal step size of ρ lies at the intersection of these two values:

[0210] .

[0211] It should be noted that the reason this embodiment requires finding the optimal solution is... The aim is to make the iteration of the augmented Lagrange multiplier method (ALM) faster here, thus accelerating the process of formula... The iteration can find the optimal sparsity coefficients in the fewest number of iterations. After finding the optimal one, the corresponding optimal sparse dictionary in the iteration process is also found. Of course, there are still some useless atoms and noise atoms in the optimal sparse dictionary that need to be removed.

[0212] In step S4, useless atoms in the optimal sparse dictionary (learning dictionary) are removed, including:

[0213] From a data reconstruction perspective, the reconstructed signal suffers from the contradiction between feature factors and noise factors, a phenomenon known as source crossover between defective feature atoms and noise atoms, or atom duality. Dictionary masking aims to artificially control the selection of important feature atoms from the data using a complete dictionary, thereby enabling the latent representation to effectively recover the original signal features. By analyzing the sparse coefficient representation, given a sparse representation method and dataset, it's impossible to know how many atoms are appropriate; a widely adopted approach is to set the number of atoms empirically. If the number is too small, the signal may not be sparse enough in the dictionary; if it's too large, more unimportant atoms will lead to a heavy computational burden. If the number of atoms can be reduced while maintaining a sparse representation, the efficiency of sparse decomposition and reconstruction can be improved. This requires reducing the number of atoms while minimizing reconstruction errors.

[0214] The learning dictionary in this embodiment is an overcomplete dictionary. It uses a dictionary mask to remove useless atoms from the overcomplete dictionary and then selects useful atoms.

[0215] For the existing training data , For the i-th sample, given an overcomplete dictionary have:

[0216] ;

[0217] in, To find the sparsity coefficients of the corresponding atoms in the overcomplete dictionary, Let be the ith sparse coefficient corresponding to the overcomplete dictionary. A set of mask vectors, If the i-th atom is a useful atom, it needs to be retained. If the i-th atom is a useless atom, it needs to be removed. Therefore, after passing through the mask vector... The transformed sparsity coefficients are expressed as follows:

[0218] ;

[0219] Simultaneously, the following conditions must be met:

[0220] ;

[0221] in, For the mask vector The converted coefficients, For Hadama accumulation, For splicing operations, For the union, For the first to the Mth samples.

[0222] In this embodiment, atoms that are not removed are identified as important atoms. Sparse decomposition is the process of selecting atoms to represent the signal, focusing on the criteria for determining the importance of atoms. In sparse representation, if a weight is non-zero, it indicates that the corresponding atom has been selected and used. Furthermore, since the atoms in the dictionary are normalized, the weight value reflects the contribution of the corresponding atom. In other words, the larger the sparse decomposition value, the greater the energy contribution of the corresponding atom. Therefore, if an atom is not frequently used in a large number of sample representations, this atom is not important and can be deleted. It is worth noting that, based on experimental results, the number of atoms cannot be reduced during the initial dictionary setup, because this would reduce the diversity of atoms, thus failing to filter out useful atoms. This embodiment aims to improve efficiency at the cost of negligible reconstruction errors.

[0223] Even after applying dictionary masks, some atoms may still exist that contribute little to the sparse representation of the signal but have not yet been removed. These atoms may be retained due to their inherent properties, but are not actually needed when reconstructing the signal. Gradient Screening (GS) provides a more refined method to identify these atoms, as follows:

[0224] Gradient screening (GS): Atoms that are never used or are infrequently used indicate a low average contribution to the signal representation. Negligible atoms can be progressively removed by sorting the absolute values ​​of their corresponding coefficients.

[0225] It should be noted that the gradient screening GS process does not delete all atoms with small coefficients at once, but rather proceeds gradually. Specifically, at the beginning, some atoms with the smallest absolute coefficient values ​​are deleted, and then the impact on signal reconstruction is observed. If the quality of the reconstructed signal does not decrease significantly, then the next batch of atoms with the smallest coefficients can be deleted.

[0226] In step S4, noise atoms in the optimal sparse dictionary (learned dictionary) are removed, including:

[0227] During the vibration process, energy is gradually absorbed into the chip solder joints. The frequency and energy distribution are closely related to the state of the solder balls between the chip substrate. A feature atom screening strategy based on energy entropy distribution is proposed:

[0228] Entropy is a complexity metric related to the randomness / uncertainty of a data sequence; simply put, entropy is a function of the probability distribution. Energy entropy is an extension of entropy within the energy domain, and it relates to the distribution of energy components. Valid atomic components occupy the majority of energy, while spurious atomic components occupy a smaller proportion. Treating each atom in the learning dictionary as an atomic component, the signal energy of the i-th atomic component is calculated as follows:

[0229] ;

[0230] The signal energy of each atomic component is normalized using the following formula:

[0231] ;

[0232] in, Let i be the signal energy of the i-th atomic component, i(t) be the t-th atomic component, and n be the number of all atomic components in the learning dictionary.

[0233] Calculate the energy entropy increment for the signal energy of each atomic component:

[0234] ;

[0235] If the value of ΔQ(i) is higher than the preset discrimination threshold, it indicates that the atomic component is more authentic, and the atomic component is retained; if the value of ΔQ(i) is lower than the preset discrimination threshold, it indicates that the atomic component is less authentic, and it is treated as a noise atom and discarded. To prevent effective signals from being excessively discarded, the discrimination threshold is set to ΔQ(i) / μ≥0.85, where μ=0.2.

[0236] Example 2

[0237] This embodiment provides a flip-chip vibration signal denoising system, including:

[0238] Acquisition module: Used to acquire vibration signals from flip chips used for defect detection;

[0239] Construction module: used to construct an initial sparse representation of the vibration signal, construct a sparse dictionary in the initial sparse representation based on the vibration signal, and construct a sparse representation model based on the sparse dictionary and the noise variance of the vibration signal.

[0240] Solving module: used to solve for the optimal sparse coefficients in the sparse representation model with respect to the sparse dictionary, and simultaneously obtain the optimal sparse dictionary corresponding to the optimal sparse coefficients;

[0241] Elimination module: used to remove useless and noisy atoms from the optimal sparse dictionary, resulting in a denoised optimal sparse dictionary;

[0242] Reconstruction module: Used to reconstruct the noisy vibration signal using the denoised optimal sparse dictionary and its corresponding optimal sparse coefficients, so as to obtain the denoised vibration signal.

[0243] Example 3

[0244] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the flip chip vibration signal denoising method described in Embodiment 1.

[0245] Example 4

[0246] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the flip-chip vibration signal denoising method described in Embodiment 1.

[0247] Example 5

[0248] This embodiment provides a computer program product, including a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the flip chip vibration signal denoising method described in Embodiment 1.

[0249] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0250] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0251] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0252] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0253] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0254] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for denoising vibration signals from a flip chip, characterized in that: include: Step S1: Acquire vibration signals from the flip chip used for defect detection; Step S2: Construct an initial sparse representation of the vibration signal, and construct a sparse dictionary in the initial sparse representation based on the vibration signal. Construct a sparse representation model based on the sparse dictionary and the noise variance of the vibration signal. The method for constructing the sparse representation model in step S2 includes: The vibration signal is sparsely reconstructed to obtain an initial sparse representation: ； Where D is the sparse dictionary, S is the vibration signal matrix, N is the noise matrix, and X is the sparse coefficient of the vibration signal Y in the sparse dictionary D. This is an acceptable error term; The initial sparse representation is transformed into an initial objective function, expressed as: ； in, To find the minimum value of the objective function, It is the l2 norm. The original signal matrix, The coefficient of the regularization term, It is a regular term; Constructing a sparse dictionary D using SVD decomposition: By applying SVD adaptive learning to the obtained vibration signal matrix S, it can be represented as: ； in, It is a left singular orthogonal matrix. It is a real matrix. for The number of feature vectors, for The i-th eigenvector, It is a right singular orthogonal matrix. for The number of feature vectors, for The i-th feature vector For singular value matrices with sparse properties, Singular value matrix The i-th singular value; left singular orthogonal matrix As a sparse dictionary D, since the sparse dictionary D is adaptively learned from the vibration signal matrix S, it is used to effectively represent feature information; The noise variance of the vibration signal matrix S is estimated using the median absolute deviation (MAD). ； in, For noise variance, This represents the median. To take the absolute value, c1 is the wavelet coefficient at the highest frequency scale layer of the wavelet decomposition; Based on noise variance and singular value matrix Construct an adaptive dual weight matrix W, and set each element in the adaptive dual weight matrix W to: ； in, Singular value matrix The i-th singular value, i∈[1,2,...,M]; Let represent the noise variance of the j-th column of the vibration signal matrix, where j∈[1,2,....,N]; The adaptive dual weight matrix W is introduced into the initial objective function and used as a sparse representation model: ； in, For Hadama accumulation, It is the l1 norm; Step S3: Solve for the optimal sparse coefficients in the sparse representation model with respect to the sparse dictionary, and simultaneously obtain the optimal sparse dictionary corresponding to the optimal sparse coefficients; The method for solving the optimal sparse coefficients of the sparse dictionary in the sparse representation model in step S3 includes: The sparse representation model is represented using the alternating direction multiplier method: ； in , , The constraint on X in the sparse representation; Since the adaptive dual weight matrix W is a positive definite matrix that varies according to the vibration signal matrix S, and has no direct relationship with the sparse coefficients, we extract it to obtain: ； in, This incorporates regularization terms that combine the dual weight matrix W; To solve It can be expressed using the augmented Lagrange function, as shown in the formula: ； Where u represents the dual variable, ρ represents the transpose of the dual variable u; ρ represents the penalty parameter, which serves as the step size in the alternating direction multiplier method solution process. To obtain the optimal Using the augmented Lagrange multiplier method (ALM) to refine the formula The iteration is as follows: ； ； ； The corresponding iterative solution is: ； ； ； in, X represents the sparse coefficients in the (k+1)th stage. Z represents the constraint variable Z of the sparse coefficients X in the k-th stage. Let u represent the dual variable in the k-th stage. Represents the solution in soft-threshold form. This represents the transpose of variable u in the k-th stage. Represents the identity matrix. Let X represent the sparse coefficients of the k-th stage. Abbreviated as And used as a soft threshold operator, the formula is: ； in, Indicates the first generation value. Indicates the second-generation value; Based on the soft threshold operator The definition will The solution can be written in the following form: ； First scenario: When The solution satisfies At that time, bring it in From To study the convergence of the algorithm at this point, we will... or Eliminate; remove Iterative and Substitute it into this moment From: ； Assume Z* is The fixed point in the solution process, i.e. Then dual error At this point, we have: ； This represents the k-th stage of the dual error. express The dual error converges if and only if the spectral radius of matrix E is less than 1. To represent the eigenvalues ​​of matrix E, let... for The eigenvalues ​​of matrix E are then expressed as: ； This represents the i-th eigenvalue. Representing the eigenvalue representation of matrix E, when When the eigenvalue is greater than 0, the above equation is less than 1, and the algorithm guarantees convergence; therefore, solving for the optimal step size is equivalent to solving the following problem: ； because about It is monotonically decreasing, simplifying the above equation to: ； D represents T The smallest eigenvalue of D; Second scenario: When The solution satisfies Then we get: ； Assume u* is The fixed point in the solution process, i.e. Then dual error At this point, we have: ； The algorithm converges at this point if and only if the spectral radius of matrix E is less than 1, and the eigenvalues ​​of matrix E are: ； when When the eigenvalue is greater than 0, the above expression is less than 1, and the algorithm guarantees convergence. about It is monotonically increasing, at this time: ； D represents T The largest eigenvalue of D; Third scenario: When Z k+1 The solution satisfies Then: ； At this time, Z k+1 The matrix E is the same as in the first case, therefore we get: ； exist The iterative process of the three cases satisfies: when When, the optimal step size ; when When, the optimal step size ; Will and This can be summarized in the following expression: ； and The corresponding spectral radius is expressed as: ； In the entire domain, the left half about Monotonically decreasing, right half about Monotonically increasing, therefore, assume If the optimal step size is located at the intersection of the two, then we get: ； Based on the obtained optimal step size Used to accelerate the augmented Lagrange multiplier method (ALM) for formulas Iteration is then performed to obtain the optimal sparsity coefficients. ; Step S4: Remove useless and noise atoms from the optimal sparse dictionary to obtain the denoised optimal sparse dictionary; Step S5: Reconstruct the noisy vibration signal using the denoised optimal sparse dictionary and its corresponding optimal sparse coefficients to obtain the denoised vibration signal.

2. The method for denoising vibration signals of flip-chip according to claim 1, characterized in that: The method for removing useless atoms from the optimal sparse dictionary in step S4 includes: The optimal sparse dictionary is used as the learning dictionary, which is an overcomplete dictionary. Useless atoms are removed from the overcomplete dictionary by using a dictionary mask, and then useful atoms are selected. To reduce the number of atoms while keeping the errors in the reconstructed and vibrational signals below a preset value, the existing training data... , For the i-th sample, given an overcomplete dictionary have: ； in, To determine the sparsity coefficients of the corresponding atoms in an overcomplete dictionary, Let i be the sparse coefficient corresponding to the overcomplete dictionary; set A set of mask vectors, If the i-th atom is a useful atom, it needs to be retained. If the i-th atom is a useless atom and needs to be removed, then it is processed through the mask vector. The transformed sparsity coefficients are expressed as: ； in, For the mask vector The converted coefficients, For Hadama accumulation, For splicing operations, For the union, For the first to the Mth samples; Simultaneously, the following conditions must be met: 。 3. The method for denoising vibration signals of flip-chip according to claim 2, characterized in that: The method of removing useless atoms from the optimal sparse dictionary in S4 further includes: updating the atoms in the overcomplete dictionary after dictionary masking, sorting the absolute values ​​of the sparsity coefficients of each atom by gradient filtering, and removing atoms that have never been used or have been used less than a preset number of times.

4. The method for denoising vibration signals of flip-chip according to claim 2, characterized in that: The method for removing noise atoms from the optimal sparse dictionary in step S4 includes: Treating each atom in the learning dictionary as an atomic component, the signal energy of the i-th atomic component in the learning dictionary is calculated using the following formula: ； The signal energy of each atomic component is normalized using the following formula: ； in, Let i be the signal energy of the i-th atomic component, i(t) be the t-th atomic component, and n be the number of all atomic components in the learning dictionary. The formula for calculating the energy entropy increment of the signal energy of each atomic component is as follows: ； If the value of ΔQ(i) is higher than the preset discrimination threshold, it means that the atomic component is more authentic and the atomic component is retained; if the value of ΔQ(i) is lower than the preset discrimination threshold, it means that the atomic component is less authentic and the atomic component is treated as noise atom and removed.

5. A flip-chip vibration signal denoising system, used to implement the flip-chip vibration signal denoising method as described in any one of claims 1 to 4, characterized in that: include: Acquisition module: Used to acquire vibration signals from flip chips used for defect detection; Construction module: used to construct an initial sparse representation of the vibration signal, construct a sparse dictionary in the initial sparse representation based on the vibration signal, and construct a sparse representation model based on the sparse dictionary and the noise variance of the vibration signal. Solving module: used to solve for the optimal sparse coefficients in the sparse representation model with respect to the sparse dictionary, and simultaneously obtain the optimal sparse dictionary corresponding to the optimal sparse coefficients; Elimination module: used to remove useless and noisy atoms from the optimal sparse dictionary, resulting in a denoised optimal sparse dictionary; Reconstruction module: used to reconstruct the noisy vibration signal using the denoised optimal sparse dictionary and its corresponding optimal sparse coefficients, so as to obtain the denoised vibration signal.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the flip chip vibration signal denoising method as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the flip chip vibration signal denoising method as described in any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the flip chip vibration signal denoising method according to any one of claims 1 to 4.

Images