Solder ball quality detection method and device based on machine learning and storage medium

Abstract

The invention belongs to the field of data processing, and relates to a solder ball quality detection method and device based on machine learning, computer equipment and a storage medium, and the method comprises the steps: obtaining solder ball waveform data, and carrying out the normalization, filtering and time window alignment processing of the solder ball waveform data, and obtaining standard solder ball waveform data; performing fractional order wavelet transform on the standard solder ball waveform data to obtain a multi-scale solder ball wavelet coefficient; performing adaptive order selection on the multi-scale solder ball wavelet coefficient to obtain a solder ball fractional order parameter; performing feature extraction and feature matrix construction on the solder ball fractional order parameters to obtain a multi-scale solder ball feature matrix; performing weighted fusion based on the multi-scale solder ball feature matrix to obtain a solder ball fusion feature vector; and inputting the solder ball fusion feature vector into a pre-trained quality classification model to obtain a solder ball quality detection result based on machine learning. According to the invention, accurate detection and classification of the quality of the solder balls can be realized.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, computer equipment, and storage medium for solder ball quality inspection based on machine learning. Background Technology

[0002] In semiconductor wire bonding, the EFO (Extended Air Formation) device is used to melt the end of the metal wire to form a free air ball (FAB), i.e., a solder ball. The size, shape, surface oxidation contamination, and hardness of the solder ball directly determine the bonding quality and reliability of the subsequent ball-bonding joint. Therefore, obtaining smooth solder balls with good sphericity, uniform size, and no oxidation during the solder ball formation process is an important condition for ensuring excellent solder joint performance.

[0003] Existing technologies typically employ a "Ball Form Monitor (BFM)" function, which judges solder ball quality by monitoring simple thresholds in the EFO discharge voltage / current waveform. If the waveform characteristics do not meet the preset threshold range, the solder ball is considered defective. However, this method of judging based on setting fixed thresholds using human experience has significant shortcomings: the threshold needs repeated adjustment when conditions such as different machines, wire materials, or shielding gases change, making it difficult to adapt and prone to misjudgments or missed detections. In addition, traditional BFM can only provide limited "pass / fail" indications and cannot fully utilize complete waveform information to identify specific defect types (such as contaminated balls, short-tailed balls, etc.). These shortcomings limit the reliability of quality monitoring in the wire bonding process, necessitating the introduction of more intelligent solder ball quality judgment methods to improve production yield. Summary of the Invention

[0004] The purpose of this application is to propose a method, device, computer equipment, and storage medium for solder ball quality inspection based on machine learning, so as to solve the problem of difficulty in accurately classifying and judging the quality of solder balls.

[0005] To address the aforementioned technical problems, this application provides a machine learning-based solder ball quality inspection method, employing the following technical solution:

[0006] Acquire solder ball waveform data, and perform normalization, filtering, and time window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data;

[0007] Perform fractional wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients;

[0008] Adaptive order selection is performed on the multi-scale solder ball wavelet coefficients to obtain the solder ball fractional-order parameters;

[0009] Feature extraction and feature matrix construction are performed on the fractional-order parameters of the solder balls to obtain a multi-scale solder ball feature matrix;

[0010] Based on the multi-scale solder ball feature matrix, a weighted fusion is performed to obtain the solder ball fusion feature vector;

[0011] The fused feature vector of the solder ball is input into a pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning.

[0012] Furthermore, the step of performing fractional wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients specifically includes:

[0013] Obtain a preset fractional wavelet basis function, and perform a discrete Fourier transform on the standard solder ball waveform data and the preset fractional wavelet basis function to obtain first frequency domain data and second frequency domain data;

[0014] Perform frequency domain fractional order calculation on the first frequency domain data to obtain frequency domain signal data;

[0015] The second frequency domain data is subjected to frequency domain fractional order calculation to obtain the basis function data;

[0016] The multi-scale solder ball wavelet coefficients are obtained by performing inverse discrete Fourier transform on the frequency domain signal data and the basis function data.

[0017] Furthermore, the step of adaptively selecting the order of the multi-scale solder ball wavelet coefficients to obtain the fractional-order parameters of the solder balls specifically includes:

[0018] Energy calculations are performed on the wavelet coefficients of the multi-scale solder balls to obtain scale energy information;

[0019] Obtain the preset fractional order range information, and construct a relational model based on the scale energy information and the fractional order range information to obtain a mapping relational model;

[0020] Based on the mapping relationship model, the objective function is defined to obtain an effective objective function;

[0021] The fractional-order parameters of the solder balls are obtained by optimizing the solution based on the effective objective function.

[0022] Furthermore, the step of extracting features and constructing feature matrices from the fractional-order parameters of the solder balls to obtain a multi-scale solder ball feature matrix specifically includes:

[0023] Single-scale feature extraction is performed on the fractional-order parameters of the solder ball to obtain the solder ball scale feature vector;

[0024] A multi-scale feature set is obtained by constructing a scale feature set based on the solder ball scale feature vector.

[0025] The feature matrix is ​​constructed based on the multi-scale feature set to obtain the multi-scale solder ball feature matrix.

[0026] Furthermore, the step of weighted fusion based on the multi-scale solder ball feature matrix to obtain the solder ball fusion feature vector specifically includes:

[0027] Information entropy will be calculated on the multi-scale solder ball feature matrix to obtain multi-scale information entropy, and weights will be calculated based on the multi-scale information entropy to obtain information entropy weight information.

[0028] The correlation between the multi-scale solder ball feature matrix and the preset quality level information will be calculated to obtain the correlation coefficient, and the weight will be calculated based on the correlation coefficient to obtain the correlation weight information.

[0029] Based on the information entropy weight information and the relevance weight information, a comprehensive weight calculation is performed to obtain effective weight information;

[0030] The solder ball fusion feature vector is obtained by performing weighted fusion calculation based on the multi-scale solder ball feature matrix and the effective weight information.

[0031] Furthermore, before the step of inputting the fused feature vector of the solder balls into the pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning, the following steps are also included:

[0032] Obtain the label data extraction identifier, and extract the label solder ball waveform dataset from the database based on the label data extraction identifier;

[0033] The labeled solder ball waveform dataset is preprocessed and the dataset is divided into standard training dataset and standard validation dataset.

[0034] The standard training dataset is input into a pre-built neural network model framework for supervised learning training to obtain an initial quality classification model;

[0035] The initial quality classification model is adjusted and optimized based on the standard validation dataset until the adjusted and optimized initial quality classification model meets the preset model quality conditions, and the initial quality classification model that meets the preset model quality conditions is taken as the quality classification model.

[0036] Furthermore, the step of inputting the fused feature vector of the solder balls into a pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning specifically includes:

[0037] Based on the convolutional layer, convolutional feature extraction is performed on the fused feature vector of the solder ball to obtain the convolutional feature of the solder ball.

[0038] The convolutional features of the solder balls are expanded by pooling based on the pooling layer to obtain a one-dimensional vector of solder ball features.

[0039] The one-dimensional vector of solder ball features is transformed by a fully connected layer to obtain the solder ball quality detection result based on machine learning.

[0040] To address the aforementioned technical problems, this application also provides a solder ball quality inspection device based on machine learning, employing the following technical solution:

[0041] The data processing module is used to acquire solder ball waveform data and perform normalization, filtering, and time window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data.

[0042] The data transformation module is used to perform fractional wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients.

[0043] The order selection module is used to adaptively select the order of the multi-scale solder ball wavelet coefficients to obtain the solder ball fractional-order parameters.

[0044] The matrix construction module is used to extract features and construct feature matrices from the fractional-order parameters of the solder balls to obtain a multi-scale solder ball feature matrix.

[0045] The weighted fusion module is used to perform weighted fusion based on the multi-scale solder ball feature matrix to obtain a solder ball fusion feature vector.

[0046] The result acquisition module is used to input the fused feature vector of the solder ball into a pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning.

[0047] To address the aforementioned technical problems, this application also provides a computer device that employs the following technical solution:

[0048] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described machine learning-based solder ball quality inspection method.

[0049] To address the aforementioned technical problems, this application also provides a computer-readable storage medium, employing the technical solution described below:

[0050] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned machine learning-based solder ball quality inspection method.

[0051] Compared with the prior art, the embodiments of this application have the following main advantages:

[0052] This application obtains solder ball waveform data and performs normalization, filtering, and time-window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data. It then performs fractional-order wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients. Adaptive order selection is applied to the multi-scale solder ball wavelet coefficients to obtain fractional-order solder ball parameters. Feature extraction and feature matrix construction are performed on the fractional-order solder ball parameters to obtain a multi-scale solder ball feature matrix. Weighted fusion is then performed based on the multi-scale solder ball feature matrix to obtain a fused solder ball feature vector. This fused feature vector is input into a pre-trained quality classification model to obtain a machine learning-based solder ball quality detection result. This effectively achieves accurate detection and classification of solder ball quality, facilitating subsequent processing based on solder ball quality. Attached Figure Description

[0053] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart of an embodiment of the machine learning-based solder ball quality inspection method according to this application;

[0055] Figure 2 This is a schematic diagram of one embodiment of the machine learning-based solder ball quality inspection device according to this application;

[0056] Figure 3 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0057] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0058] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0059] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0060] refer to Figure 1 The diagram illustrates a flowchart of an embodiment of a machine learning-based solder ball quality inspection method according to this application. The machine learning-based solder ball quality inspection method includes the following steps:

[0061] Step S10: Obtain solder ball waveform data, and perform normalization, filtering, and time window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data;

[0062] In this embodiment, the solder ball waveform data refers to the voltage waveform data corresponding to the solder ball formed by the EFO device melting the end of the metal wire during semiconductor wire bonding. The solder ball waveform data can be acquired via a high-speed analog-to-digital converter (ADC) after each EFO discharge. The EFO device is an electronic flame extinguishing device that uses high-voltage discharge to generate an electric arc. The high temperature generated by the arc instantly melts the end of the metal wire, forming a spherical shape. A high-speed analog-to-digital converter (ADC) is a device that can convert continuous analog signals (such as voltage signals) into discrete digital signals.

[0063] The solder ball waveform data is first normalized to scale the amplitude to the range of [-1, 1], eliminating amplitude differences caused by different devices or environments. The normalization process uses a maximum-minimum normalization method, specifically calculated as: X_norm = (X - X_min) / (X_max - X_min) * 2 - 1; where X is the original waveform data point, X_min and X_max are the minimum and maximum values ​​of the waveform data, respectively, and X_norm is the normalized data point. Next, a bandpass filter is applied to the normalized solder ball waveform data to remove high-frequency noise and low-frequency drift. The low cutoff frequency of the bandpass filter can be set to 100Hz, and the high cutoff frequency can be set to 5MHz. Finally, time window alignment is performed on the noise-filtered solder ball waveform data, aligning all waveforms on the time axis based on the trigger point formed by the solder ball. In this embodiment, the time window length is uniformly set to 2048 sampling points, with a sampling rate of 10MHz. This time window length can be adjusted according to actual conditions.

[0064] Step S20: Perform fractional wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients;

[0065] In this embodiment, fractional wavelet transform is a coefficient transformation method that considers the local characteristics of the signal in the time-frequency domain and adjusts the shape of the wavelet basis function by fractional parameters to adapt to the analysis needs of different signal characteristics. Multi-scale solder ball wavelet coefficients refer to the wavelet coefficients corresponding to standard solder ball waveform data at multiple scales. These wavelet coefficients can reflect the energy distribution and characteristics at different time scales during the solder ball formation process.

[0066] Step S30: Adaptively select the order of the multi-scale solder ball wavelet coefficients to obtain the solder ball fractional-order parameters;

[0067] In this embodiment, adaptive order selection is a method for dynamically adjusting the fractional-order parameter α. It automatically selects the optimal fractional-order parameter based on the local features of the signal to identify signal features at different scales, thereby more accurately capturing the dynamic changes during the solder ball formation process. The solder ball fractional-order parameter is the optimal fractional-order parameter in multi-scale analysis, representing the best feature extraction parameter for the solder ball waveform signal under fractional-order wavelet transform. It reflects the fractional-order calculus order that best reveals the dynamic characteristics during solder ball formation within the fractional-order wavelet transform framework.

[0068] Step S40: Perform feature extraction and feature matrix construction on the fractional-order parameters of the solder balls to obtain a multi-scale solder ball feature matrix;

[0069] In this embodiment, feature extraction and feature matrix construction of the fractional-order parameters of solder balls refer to extracting features of the fractional-order parameters of solder balls at different scales, and then using the extracted features to construct feature matrices at multiple scales. The multi-scale solder ball feature matrix is ​​a set of solder ball features at multiple scales, which is represented as a feature matrix.

[0070] Step S50: Perform weighted fusion based on the multi-scale solder ball feature matrix to obtain the solder ball fusion feature vector;

[0071] In this embodiment, weighted fusion based on the multi-scale solder ball feature matrix refers to weighting and fusing the multi-scale solder ball feature matrix according to the calculated effective weights to form a fused feature vector representing the overall solder ball feature. The effective weights are calculated by comprehensively considering the weights obtained from calculating the information entropy and correlation of the multi-scale solder ball feature matrix. These effective weights take into account the information richness and redundancy of the features, thereby improving the quality of the fused features.

[0072] Step S60: Input the fused feature vector of the solder ball into the pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning.

[0073] In this embodiment, the pre-trained quality classification model can employ a neural network model or a support vector machine algorithm. Before inputting the solder ball fusion feature vector into the pre-trained quality classification model, the solder ball fusion data feature vector needs to be standardized to obtain a standard fusion feature vector that is convenient for the model to calculate. After obtaining the standard fusion feature vector, it is then used as model input and category prediction to obtain the corresponding machine learning-based solder ball quality detection result.

[0074] This application obtains solder ball waveform data and performs normalization, filtering, and time-window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data. It then performs fractional-order wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients. Adaptive order selection is applied to the multi-scale solder ball wavelet coefficients to obtain fractional-order solder ball parameters. Feature extraction and feature matrix construction are performed on the fractional-order solder ball parameters to obtain a multi-scale solder ball feature matrix. Weighted fusion is then performed based on the multi-scale solder ball feature matrix to obtain a fused solder ball feature vector. This fused feature vector is input into a pre-trained quality classification model to obtain a machine learning-based solder ball quality detection result. This effectively achieves accurate detection and classification of solder ball quality, facilitating subsequent processing based on solder ball quality.

[0075] In some optional implementations of this embodiment, performing fractional wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients includes the following steps:

[0076] Obtain a preset fractional wavelet basis function, and perform a discrete Fourier transform on the standard solder ball waveform data and the preset fractional wavelet basis function to obtain first frequency domain data and second frequency domain data;

[0077] In this embodiment, the preset fractional-order wavelet basis function is selected as the fractional-order Meyer wavelet, and the fractional-order parameter is initially set to 0.5, representing the fractional order of the wavelet basis function. Discrete Fourier transforms are performed on the standard solder ball waveform data x(t) and the preset fractional-order wavelet basis function ψ(t) according to the following formulas: X(ω) = FFT[x(t)]; Ψ(ω) = FFT[ψ(t)]; where FFT represents the Fast Fourier Transform algorithm, X(ω) is the first frequency domain data, and Ψ(ω) is the second frequency domain data.

[0078] Perform frequency domain fractional order calculation on the first frequency domain data to obtain frequency domain signal data;

[0079] In this embodiment, the frequency domain fractional order calculation of the first frequency domain data is performed according to the following formula: X_α(ω)=X(ω)·|ω|^α; where X(ω) is the first frequency domain data, α is the fractional order parameter, |ω| represents the absolute value of the frequency, and X_α(ω) is the frequency domain signal data.

[0080] The second frequency domain data is subjected to frequency domain fractional order calculation to obtain the basis function data;

[0081] In this embodiment, the fractional frequency order calculation of the second frequency domain data is performed according to the following formula: Ψ_α(ω)=Ψ(ω)·|ω|^α; where Ψ(ω) is the second frequency domain data, α is the fractional order parameter, |ω| represents the absolute value of the frequency, and Ψ_α(ω) is the basis function data.

[0082] The multi-scale solder ball wavelet coefficients are obtained by performing inverse discrete Fourier transform on the frequency domain signal data and the basis function data.

[0083] In this embodiment, the inverse discrete Fourier transform (IFFT) of the frequency domain signal data and basis function data is performed according to the following formula: W_α(a, b) = IFFT[X_α(ω)·Ψ_α*(ω)]; where IFFT represents the inverse fast Fourier transform, Ψ_α*(ω) represents the conjugate of Ψ_α(ω), W_α(a, b) are the multi-scale wavelet coefficients, a represents the scale parameter, and b represents the translation parameter. In this embodiment, by adjusting the value of the scale parameter a, wavelet coefficients at different scales can be obtained, forming multi-scale analysis results. For example, the value range of the scale parameter a is [1, 2, 4, 8, 16, 32], with a total of 6 scales, and 2048 wavelet coefficients are obtained at each scale.

[0084] This embodiment obtains a preset fractional-order wavelet basis function, performs a Discrete Fourier Transform on the standard solder ball waveform data and the preset fractional-order wavelet basis function to obtain first frequency domain data and second frequency domain data; performs a fractional-order frequency domain calculation on the first frequency domain data to obtain frequency domain signal data; performs a fractional-order frequency domain calculation on the second frequency domain data to obtain basis function data; and performs an Inverse Discrete Fourier Transform on the frequency domain signal data and the basis function data to obtain the multi-scale solder ball wavelet coefficients. This effectively obtains multi-scale solder ball wavelet coefficients based on the standard solder ball waveform, facilitating subsequent adaptive order selection processing.

[0085] In some optional implementations of this embodiment, the adaptive order selection of the multi-scale solder ball wavelet coefficients to obtain the solder ball fractional-order parameters includes the following steps:

[0086] Energy calculations are performed on the wavelet coefficients of the multi-scale solder balls to obtain scale energy information;

[0087] In this embodiment, scale energy information refers to the energy value corresponding to the wavelet coefficients of the solder ball at each scale. The energy calculation of the wavelet coefficients at each scale is performed according to the following formula: E(a)=Σ|W_α(a,b)|^2; where E(a) represents the energy value at scale a, and Σ represents the summation over all translation parameters b.

[0088] Obtain the preset fractional order range information, and construct a relational model based on the scale energy information and the fractional order range information to obtain a mapping relational model;

[0089] In this embodiment, the preset fractional order range information can be set to [0.1, 0.9], where the step size is 0.1, and there are a total of 9 candidate order values. Based on the scale energy information and the fractional order range information, the mapping relationship model can be constructed according to the following formula: M(α) = f(E(a), α); where M(α) represents the mapping relationship model, f represents the mapping function, and α represents the fractional order. The mapping function can be obtained using a polynomial fitting method, expressing the relationship between scale energy and fractional order as a third-order polynomial: f(E(a), α) = c0 + c1·E(a) + c2·E(a) 2 +c3·E(a) 3 Where c0, c1, c2, and c3 are polynomial coefficients, determined by the least squares method.

[0090] Based on the mapping relationship model, the objective function is defined to obtain an effective objective function;

[0091] In this embodiment, the objective function is defined according to the following formula: J(α)=||M(α)-M_target|| 2 +λ·R(α); where M_target represents the target mapping value, usually set to the mapping value of the ideal solder ball; λ represents the regularization coefficient, set to 0.01; R(α) represents the regularization term, used to prevent overfitting.

[0092] The fractional-order parameters of the solder balls are obtained by optimizing the solution based on the effective objective function.

[0093] In this embodiment, the objective function can be optimized using gradient descent to obtain the optimal solder ball fractional-order parameters. In this embodiment, the number of iterations can be set to 100, the learning rate to 0.01, and the convergence threshold to 1e-6.

[0094] This embodiment calculates the energy of the multi-scale solder ball wavelet coefficients to obtain scale energy information; acquires a preset fractional order range; constructs a relational model based on the scale energy information and the fractional order range information to obtain a mapping relational model; defines an objective function based on the mapping relational model to obtain an effective objective function; and optimizes the solution based on the effective objective function to obtain the fractional order parameters of the solder balls. This effectively achieves adaptive order selection based on the multi-scale solder ball wavelet coefficients to obtain reliable fractional order parameters for the solder balls, facilitating the subsequent acquisition of the multi-scale solder ball feature matrix.

[0095] In some optional implementations of this embodiment, the step of extracting features and constructing a feature matrix from the fractional-order parameters of the solder balls to obtain a multi-scale solder ball feature matrix includes the following steps:

[0096] Single-scale feature extraction is performed on the fractional-order parameters of the solder ball to obtain the solder ball scale feature vector;

[0097] In this embodiment, single-scale feature extraction of the solder ball fractional-order parameters refers to extracting preset features for the solder ball fractional-order parameters at each scale. The features corresponding to single-scale feature extraction include: mean: μ=(1 / N)·Σα_i; standard deviation: σ=sqrt((1 / N)·Σ(α_i-μ)). 2 ); kurtosis: K=(1 / N)·Σ((α_i-μ) / σ) 4 ; Skewness: S=(1 / N)·Σ((α_i-μ) / σ) 3 Maximum value: Max = max(α_i); Minimum value: Min = min(α_i); Energy: E = Σ|α_i| 2 Entropy: H = -Σp(α_i)·log(p(α_i)); where N represents the number of parameters, α_i represents the i-th fractional parameter, and p(α_i) represents the probability distribution of α_i.

[0098] A multi-scale feature set is obtained by constructing a scale feature set based on the solder ball scale feature vector.

[0099] In this embodiment, the construction of the scale feature set refers to combining the solder ball scale feature vectors at each scale to form a multi-scale feature set. For example, the solder ball scale feature vectors at different scales are represented by f. s1 f s2 , ..., f sx Then the multi-scale feature set is represented as F = {f s1 f s2 , ..., f sx}

[0100] The feature matrix is ​​constructed based on the multi-scale feature set to obtain the multi-scale solder ball feature matrix.

[0101] In this embodiment, a multi-scale solder ball feature matrix is ​​constructed by arranging the feature vectors in the multi-scale feature set in scale order.

[0102] This embodiment extracts single-scale features from the fractional-order parameters of the solder balls to obtain a solder ball scale feature vector; constructs a scale feature set based on the solder ball scale feature vector to obtain a multi-scale feature set; and constructs a feature matrix based on the multi-scale feature set to obtain the multi-scale solder ball feature matrix. This effectively achieves the acquisition of the corresponding multi-scale solder ball feature matrix based on the fractional-order parameters of the solder balls, facilitating subsequent weighted fusion processing.

[0103] In some optional implementations of this embodiment, the step of weighted fusion based on the multi-scale solder ball feature matrix to obtain the solder ball fused feature vector includes the following steps:

[0104] Information entropy will be calculated on the multi-scale solder ball feature matrix to obtain multi-scale information entropy, and weights will be calculated based on the multi-scale information entropy to obtain information entropy weight information.

[0105] In this embodiment, the information entropy of the multi-scale solder ball feature matrix is ​​calculated according to the following formula: H_j=-Σ(p_ij·log(p_ij)); where H_j represents the information entropy of the j-th feature, and p_ij represents the normalized value of the j-th feature at the i-th scale. The weights are calculated based on the information entropy according to the following formula: w_j^H=(1-H_j) / Σ(1-H_j); where w_j^H represents the information entropy weight of the j-th feature.

[0106] The correlation between the multi-scale solder ball feature matrix and the preset quality level information will be calculated to obtain the correlation coefficient, and the weight will be calculated based on the correlation coefficient to obtain the correlation weight information.

[0107] In this embodiment, the preset quality level information includes three levels: "normal solder balls," "defective solder balls," and "short-tailed solder balls," represented by the values ​​3, 2, and 1, respectively. The correlation coefficient between the multi-scale solder ball feature matrix and the quality level is calculated using the following formula: r_j = |corr(F_j, Q)|; where r_j represents the correlation coefficient between the j-th feature and the quality level Q, F_j represents the vector of the j-th feature, and corr represents the correlation calculation function, using the Pearson correlation coefficient. The weights are calculated based on the correlation coefficients using the following formula: w_j^R = r_j / Σr_j; where w_j^R represents the correlation weight of the j-th feature.

[0108] Based on the information entropy weight information and the relevance weight information, a comprehensive weight calculation is performed to obtain effective weight information;

[0109] In this embodiment, the effective weight is calculated by combining the information entropy weight and the relevance weight according to the following formula: w_j=β·w_j^H+(1-β)·w_j^R; where β is the balance coefficient, set to 0.5, indicating that the information entropy weight and the relevance weight are of equal importance.

[0110] The solder ball fusion feature vector is obtained by performing weighted fusion calculation based on the multi-scale solder ball feature matrix and the effective weight information.

[0111] In this embodiment, weighted fusion based on the multi-scale solder ball feature matrix and effective weights is performed according to the following formula: F = ΣΣ(F_ij·w_j); where F represents the solder ball fusion feature vector, and F_ij represents the value of the j-th feature at the i-th scale. For example, through weighted fusion, the 6×8 matrix features are compressed into a 48-dimensional feature vector, i.e., the solder ball fusion feature vector.

[0112] This embodiment calculates the information entropy of the multi-scale solder ball feature matrix to obtain multi-scale information entropy, and then calculates weights based on the multi-scale information entropy to obtain information entropy weight information. It also calculates the correlation between the multi-scale solder ball feature matrix and preset quality level information to obtain a correlation coefficient, and then calculates weights based on the correlation coefficient to obtain correlation weight information. Finally, it performs a comprehensive weight calculation based on the information entropy weight information and the correlation weight information to obtain effective weight information. Finally, it performs a weighted fusion calculation based on the multi-scale solder ball feature matrix and the effective weight information to obtain the solder ball fusion feature vector. This effectively achieves the acquisition of the solder ball fusion feature vector corresponding to the multi-scale solder ball feature matrix, facilitating subsequent model prediction processing of solder ball quality.

[0113] In some optional implementations of this embodiment, before the step of inputting the solder ball fusion feature vector into a pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning, the following steps are also included:

[0114] Obtain the label data extraction identifier, and extract the label solder ball waveform dataset from the database based on the label data extraction identifier;

[0115] In this embodiment, the tag data extraction identifier is a unique identifier for the corresponding tag solder ball waveform dataset. This identifier can be a dataset ID or a UUID. The system can traverse and query the database based on the tag data extraction identifier to extract the corresponding tag solder ball waveform dataset. The tag solder ball waveform dataset is a dataset that labels historical solder ball waveform data extracted from historical data. The historical solder ball waveform data refers to historically acquired solder ball waveform data. This acquired solder ball waveform data is the voltage waveform data acquired by the system through a high-speed analog-to-digital converter (ADC) during the discharge gap after the EFO discharge phase of each bonding cycle. For example, in a specific embodiment, the sampling rate of the high-speed ADC is set to 10MHz, and the waveform length is 2048 points to cover the voltage changes throughout the discharge process. The discrete voltage time-series data acquired by the high-speed ADC is cached in array form to form the acquired solder ball waveform data. Tagging refers to adding quality classification labels to the acquired solder ball waveform data to form data tags. This tagging step can be performed automatically based on a machine learning model.

[0116] The labeled solder ball waveform dataset is preprocessed and the dataset is divided into standard training dataset and standard validation dataset.

[0117] In this embodiment, the preprocessing of the labeled solder ball waveform dataset includes normalization: scaling the waveform amplitude to a fixed range (e.g., [-1, 1] or [0, 1]) to eliminate the impact of amplitude differences caused by different devices or environments on model judgment; filtering and denoising: digital filters (such as bandpass filters) can be used as needed to remove high-frequency noise or low-frequency drift. For example, a high-pass filter with a cutoff frequency higher than the main frequency band of the EFO spark can be applied to filter out the residual DC component after discharge, and a low-pass filter can be applied to suppress ADC quantization noise. The filtered waveform more prominently reflects the physical process characteristics of solder ball formation; window truncation or alignment: when the acquired waveform contains redundant silent intervals, sub-waveforms of the main discharge interval can be truncated. In addition, to align all waveforms on the time axis, the waveform starting points can be aligned based on the spark trigger point. Dataset partitioning refers to dividing the preprocessed standard labeled solder ball waveform dataset into different standard training and standard validation datasets. These datasets consist of different proportions of the preprocessed standard labeled solder ball waveform dataset; for example, the standard training dataset might comprise 70% of the standard labeled solder ball waveform dataset, while the standard validation dataset might comprise 30%. The proportions of the standard training and validation datasets can be adjusted according to specific needs.

[0118] The standard training dataset is input into a pre-built neural network model framework for supervised learning training to obtain an initial quality classification model;

[0119] In this embodiment, the pre-built neural network model framework can adopt a Convolutional Neural Network (CNN) model. The structure and parameters of the CNN model can be set according to actual needs. In this embodiment, the CNN model can adopt an architecture of two one-dimensional convolutional layers, two pooling layers, and then one or two fully connected layers, finally outputting the classification result. The hyperparameters such as the length N and number of convolutional kernels in each layer, and the size of the pooling window can be determined through offline optimization experiments to control computational overhead while ensuring model accuracy. During the training phase, the CNN model is supervised learning using a standard training dataset for quality classification, allowing it to adjust its internal weights to recognize various waveform patterns. During the adjustment phase, the initially trained CNN model is validated and adjusted using a standard validation dataset. During the inference phase, the trained model is deployed to the field equipment to perform forward computation on newly acquired waveforms in real time and output the confidence score of each quality category.

[0120] The initial quality classification model is adjusted and optimized based on the standard validation dataset until the adjusted and optimized initial quality classification model meets the preset model quality conditions, and the initial quality classification model that meets the preset model quality conditions is taken as the quality classification model.

[0121] In this embodiment, evaluating and optimizing the initial quality classification model can be achieved by adjusting the model's hyperparameters (such as learning rate, regularization parameters, etc.), modifying the model structure (such as increasing or decreasing the number of network layers, adjusting the number of neurons, etc.), and using data augmentation techniques (such as performing random translations, scaling, etc., on the solder ball waveform data) to expand the training dataset. For example, if the classification accuracy on the validation set is low, the model's convergence speed may be accelerated by adjusting the learning rate, or overfitting may be prevented by adding a regularization term. Preset model quality conditions are standards for measuring whether the model meets the requirements of practical applications. These preset model quality conditions include metrics such as classification accuracy, recall, and F1 score. For example, in machine learning-based solder ball quality detection, the model may be required to achieve a classification accuracy of over 95% and a recall of over 90% on the validation set. The optimized model is evaluated using these metrics, and only when the model meets all preset quality conditions can it be used as the final quality classification model in actual production.

[0122] In this embodiment, the system makes quality judgments based on the results output by the quality classification model. If the solder ball is determined to be normal, the process continues to the next step of ball pressing and bonding operations; if the solder ball is determined to be defective (such as contamination, short tails, etc.), an equipment alarm is triggered or the current solder joint is discarded and a new solder ball is formed to prevent defective solder balls from entering the welding process. The judgment results can also be displayed in real time through the human-machine interface for operator reference. In addition, all waveforms and judgment results can be stored in the equipment database for traceability analysis and continuous model optimization.

[0123] This embodiment obtains a label data extraction identifier and extracts a labeled solder ball waveform dataset from a database based on the identifier. The labeled solder ball waveform dataset is preprocessed and split into standard training and standard validation datasets. The standard training dataset is input into a pre-built neural network model framework for supervised learning training to obtain an initial quality classification model. The initial quality classification model is then adjusted and optimized based on the standard validation dataset until it meets preset model quality conditions. This optimized initial quality classification model is then used as the final quality classification model. This allows for the training of an accurate and reliable machine learning-based solder ball quality detection model based on the labeled solder ball waveform dataset, facilitating subsequent predictive analysis of solder ball quality categories.

[0124] In some optional implementations of this embodiment, the step of inputting the solder ball fusion feature vector into a pre-trained quality classification model to obtain a machine learning-based solder ball quality detection result includes the following steps:

[0125] Based on the convolutional layer, convolutional feature extraction is performed on the fused feature vector of the solder ball to obtain the convolutional feature of the solder ball.

[0126] In this embodiment, the first convolutional layer receives a fused solder ball feature vector of length 2048 and applies several 1×N (e.g., 1×5 or 1×7) convolutional filters pointwise to generate a corresponding number of feature maps. Each convolutional filter is equivalent to a matched filter that extracts a specific pattern from the waveform, capable of detecting local features such as spikes at the moment of arc triggering and voltage drop slopes. Nonlinearity is introduced through the ReLU activation function, enabling the features to distinguish between different patterns. Subsequent convolutional layers continue to extract higher-level, more abstract solder ball convolutional features, such as combined waveform shape patterns, based on the feature maps of the previous layer.

[0127] The convolutional features of the solder balls are expanded by pooling based on the pooling layer to obtain a one-dimensional vector of solder ball features.

[0128] In this embodiment, after several convolutional layers, a max-pooling layer is inserted to downsample the solder ball convolutional features. For example, a pooling window length of 2 or 4 is used to compress the feature map length by half. This not only reduces the subsequent computation but also makes the features invariant to small positional shifts. After alternating stacking of convolutions and pooling, a small information tensor containing rich features is finally obtained. This tensor is flattened into a one-dimensional vector of solder ball features and fed into a fully connected layer.

[0129] The one-dimensional vector of solder ball features is transformed by a fully connected layer to obtain the solder ball quality detection result based on machine learning.

[0130] In this embodiment, the fully connected layer performs linear combination and nonlinear transformation on the expanded one-dimensional vector of solder ball features, fusing various local features to determine the category of the overall waveform. Typically, one or two fully connected layers are set to progressively reduce the feature dimension, with the number of output neurons in the last layer equal to the number of categories to be determined (e.g., 3 or 4 categories). The output layer uses the Softmax function to convert the output of each neuron into category probabilities. Thus, the model provides a probability estimate for each case of the input waveform, such as "normal solder ball," "contaminated solder ball," or "short-tailed solder ball." The system can output the category with the highest probability as the determination result to obtain a solder ball quality detection result based on machine learning.

[0131] This embodiment extracts solder ball convolutional features from the fused feature vector using a convolutional layer; it then expands these features using a pooling layer to obtain a one-dimensional feature vector; and finally, it transforms this one-dimensional feature vector using a fully connected layer to obtain the machine learning-based solder ball quality detection result. This effectively achieves prediction processing based on a pre-trained quality classification model, resulting in the corresponding machine learning-based solder ball quality detection result.

[0132] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0133] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0134] Further reference Figure 2 As a response to the above Figure 1 The implementation of the method shown in this application provides an embodiment of a machine learning-based solder ball quality inspection device, which is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0135] like Figure 2 As shown, the machine learning-based solder ball quality inspection device 700 described in this embodiment includes: a data processing module 701, a data transformation module 702, an order selection module 703, a matrix construction module 704, a weighted fusion module 705, and a result acquisition module 706. Wherein:

[0136] The data processing module 701 is used to acquire solder ball waveform data and perform normalization, filtering, and time window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data.

[0137] Data transformation module 702 is used to perform fractional wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients;

[0138] The order selection module 703 is used to adaptively select the order of the multi-scale solder ball wavelet coefficients to obtain the solder ball fractional-order parameters.

[0139] The matrix construction module 704 is used to extract features and construct feature matrices from the fractional-order parameters of the solder balls to obtain a multi-scale solder ball feature matrix.

[0140] The weighted fusion module 705 is used to perform weighted fusion based on the multi-scale solder ball feature matrix to obtain a solder ball fusion feature vector.

[0141] The result acquisition module 706 is used to input the fused feature vector of the solder ball into a pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning.

[0142] This embodiment employs the aforementioned machine learning-based solder ball quality detection device to acquire solder ball waveform data. The device then performs normalization, filtering, and time-window alignment on the solder ball waveform data to obtain standard solder ball waveform data. Fractional-order wavelet transform is applied to the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients. Adaptive order selection is performed on the multi-scale solder ball wavelet coefficients to obtain fractional-order solder ball parameters. Feature extraction and feature matrix construction are performed on the fractional-order solder ball parameters to obtain a multi-scale solder ball feature matrix. Weighted fusion is performed based on the multi-scale solder ball feature matrix to obtain a fused solder ball feature vector. This fused feature vector is input into a pre-trained quality classification model to obtain the machine learning-based solder ball quality detection result. This effectively achieves accurate detection and classification of solder ball quality, facilitating subsequent processing based on solder ball quality.

[0143] In some optional implementations of this embodiment, the data transformation module 702 includes: a data transformation unit, a first data calculation unit, a second data calculation unit, and a data inverse transformation unit. Wherein:

[0144] The data transformation unit is used to obtain a preset fractional wavelet basis function, and to perform a discrete Fourier transform on the standard solder ball waveform data and the preset fractional wavelet basis function to obtain first frequency domain data and second frequency domain data.

[0145] The first data calculation unit is used to perform frequency domain fractional order calculation on the first frequency domain data to obtain frequency domain signal data;

[0146] The second data calculation unit is used to perform frequency domain fractional-order calculation on the second frequency domain data to obtain basis function data;

[0147] The data inverse transform unit is used to perform inverse discrete Fourier transform on the frequency domain signal data and the basis function data to obtain the multi-scale solder ball wavelet coefficients.

[0148] This embodiment uses a data transformation module 702, which includes a data transformation unit, a first data calculation unit, a second data calculation unit, and a data inverse transformation unit, to effectively obtain multi-scale solder ball wavelet coefficients based on the standard solder ball waveform, so as to facilitate subsequent adaptive order selection processing.

[0149] In some optional implementations of this embodiment, the order selection module 703 includes: an energy calculation unit, a model building unit, a function definition unit, and a function solving unit. Wherein:

[0150] The energy calculation unit is used to perform energy calculation on the multi-scale solder ball wavelet coefficients to obtain scale energy information.

[0151] The model building unit is used to obtain preset fractional order range information, and to build a relational model based on the scale energy information and the fractional order range information to obtain a mapping relational model.

[0152] The function definition unit is used to define the objective function according to the mapping relationship model to obtain an effective objective function;

[0153] The function solving unit is used to optimize the solution based on the effective objective function to obtain the fractional-order parameters of the solder balls.

[0154] This embodiment effectively achieves adaptive order selection based on multi-scale weld ball wavelet coefficients by setting up an order selection module 703, which includes an energy calculation unit, a model building unit, a function definition unit, and a function solving unit, so as to obtain reliable weld ball fractional-order parameters and facilitate the subsequent acquisition of multi-scale weld ball feature matrices.

[0155] In some optional implementations of this embodiment, the matrix construction module 704 includes: a scale feature extraction unit, a feature set construction unit, and a feature matrix construction unit. Wherein:

[0156] The scale feature extraction unit is used to perform single-scale feature extraction on the solder ball fractional-order parameters to obtain the solder ball scale feature vector.

[0157] The feature set construction unit is used to construct a scale feature set based on the solder ball scale feature vector to obtain a multi-scale feature set.

[0158] The feature matrix construction unit is used to construct the feature matrix based on the multi-scale feature set to obtain the multi-scale solder ball feature matrix.

[0159] This embodiment effectively achieves the acquisition of the corresponding multi-scale solder ball feature matrix based on the solder ball fractional-order parameters by setting a matrix construction module 704, which includes a scale feature extraction unit, a feature set construction unit, and a feature matrix construction unit, so as to facilitate subsequent weighted fusion processing.

[0160] In some optional implementations of this embodiment, the weighted fusion module 705 includes: a first weight calculation unit, a second weight calculation unit, a comprehensive weight calculation unit, and a weighted fusion unit.

[0161] in:

[0162] The first weight calculation unit is used to calculate the information entropy of the multi-scale solder ball feature matrix to obtain the multi-scale information entropy, and to calculate the weight based on the multi-scale information entropy to obtain information entropy weight information.

[0163] The second weight calculation unit is used to perform correlation calculation on the multi-scale solder ball feature matrix and the preset quality level information to obtain the correlation coefficient, and to perform weight calculation based on the correlation coefficient to obtain the correlation weight information.

[0164] The comprehensive weight calculation unit is used to perform comprehensive weight calculation based on the information entropy weight information and the relevance weight information to obtain effective weight information.

[0165] The weighted fusion unit is used to perform weighted fusion calculation based on the multi-scale solder ball feature matrix and the effective weight information to obtain the solder ball fusion feature vector.

[0166] This embodiment effectively obtains the weld ball fusion feature vector corresponding to the multi-scale weld ball feature matrix by setting up a weighted fusion module 705 including a first weight calculation unit, a second weight calculation unit, a comprehensive weight calculation unit, and a weighted fusion unit, so as to facilitate subsequent model prediction processing of weld ball quality.

[0167] In some optional implementations of this embodiment, before the result acquisition module 706, the system further includes: a label data extraction unit, a label data processing unit, an initial model training unit, and a model adjustment and optimization unit. Wherein:

[0168] The label data extraction unit is used to obtain a label data extraction identifier and extract a label solder ball waveform dataset from the database based on the label data extraction identifier.

[0169] The tag data processing unit is used to preprocess and perform dataset segmentation on the tag solder ball waveform dataset to obtain a standard training dataset and a standard verification dataset.

[0170] The initial model training unit is used to input the standard training dataset into a pre-built neural network model framework for supervised learning training to obtain an initial quality classification model;

[0171] The model tuning and optimization unit is used to tune and optimize the initial quality classification model based on the standard validation dataset until the tuned and optimized initial quality classification model meets the preset model quality conditions, and the initial quality classification model that meets the preset model quality conditions is used as the quality classification model.

[0172] This embodiment sets up a label data extraction unit, a label data processing unit, an initial model training unit, and a model adjustment and optimization unit before the result acquisition module 706. This enables the effective training of an accurate and reliable machine learning-based solder ball quality detection model based on the label solder ball waveform dataset, facilitating subsequent predictive analysis of solder ball quality categories.

[0173] In some optional implementations of this embodiment, the result acquisition module 706 includes: a convolutional feature extraction unit, a pooling unrolling unit, and a feature output unit. Wherein:

[0174] The convolutional feature extraction unit is used to perform convolutional feature extraction on the solder ball fusion feature vector based on the convolutional layer to obtain solder ball convolutional features.

[0175] The pooling unpacking unit is used to perform pooling unpacking on the solder ball convolutional features based on the pooling layer to obtain a one-dimensional vector of solder ball features.

[0176] The feature output unit is used to perform output transformation on the one-dimensional vector of the solder ball features based on the fully connected layer to obtain the solder ball quality detection result based on machine learning.

[0177] This embodiment effectively achieves prediction processing based on a pre-trained quality classification model by setting up a result acquisition module 706 including a convolutional feature extraction unit, a pooling unrolling unit, and a feature output unit, thereby obtaining the corresponding machine learning-based solder ball quality detection result.

[0178] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 3 , Figure 3 This is a basic structural block diagram of the computer device in this embodiment.

[0179] The computer device 8 includes a memory 81, a processor 82, and a network interface 83 that are interconnected via a system bus. It should be noted that only the computer device 8 with components 81-83 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0180] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0181] The memory 81 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 81 may be an internal storage unit of the computer device 8, such as the hard disk or memory of the computer device 8. In other embodiments, the memory 81 may also be an external storage device of the computer device 8, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the computer device 8. Of course, the memory 81 may also include both the internal storage unit and its external storage device of the computer device 8. In this embodiment, the memory 81 is typically used to store the operating system and various application software installed on the computer device 8, such as the program code of a machine learning-based solder ball quality detection method. In addition, the memory 81 can also be used to temporarily store various types of data that have been output or will be output.

[0182] In some embodiments, the processor 82 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 82 is typically used to control the overall operation of the computer device 8. In this embodiment, the processor 82 is used to run program code stored in the memory 81 or process data, for example, to run the program code of the machine learning-based solder ball quality inspection method.

[0183] The network interface 83 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 8 and other electronic devices.

[0184] This embodiment, using the aforementioned computer equipment, can acquire solder ball waveform data and perform normalization, filtering, and time-window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data. Fractional wavelet transform is then performed on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients. Adaptive order selection is performed on the multi-scale solder ball wavelet coefficients to obtain fractional-order solder ball parameters. Feature extraction and feature matrix construction are then performed on the fractional-order solder ball parameters to obtain a multi-scale solder ball feature matrix. Weighted fusion is then performed based on the multi-scale solder ball feature matrix to obtain a fused solder ball feature vector. This fused feature vector is input into a pre-trained quality classification model to obtain a machine learning-based solder ball quality detection result. This effectively achieves accurate detection and classification of solder ball quality, facilitating subsequent processing based on solder ball quality.

[0185] This application also provides another embodiment, namely, providing a computer-readable storage medium storing a machine learning-based solder ball quality inspection program, which can be executed by at least one processor to perform the steps of the machine learning-based solder ball quality inspection method described above.

[0186] This embodiment, by employing the aforementioned computer-readable storage medium, can acquire solder ball waveform data and perform normalization, filtering, and time-window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data. Fractional-order wavelet transform is then performed on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients. Adaptive order selection is performed on the multi-scale solder ball wavelet coefficients to obtain fractional-order solder ball parameters. Feature extraction and feature matrix construction are then performed on the fractional-order solder ball parameters to obtain a multi-scale solder ball feature matrix. Weighted fusion is then performed based on the multi-scale solder ball feature matrix to obtain a fused solder ball feature vector. This fused feature vector is input into a pre-trained quality classification model to obtain a machine learning-based solder ball quality detection result. This effectively achieves accurate detection and classification of solder ball quality, facilitating subsequent processing based on solder ball quality.

[0187] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0188] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A method for inspecting solder ball quality based on machine learning, characterized in that, Includes the following steps: Acquire solder ball waveform data, and perform normalization, filtering, and time window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data; Perform fractional wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients; Adaptive order selection is performed on the multi-scale solder ball wavelet coefficients to obtain the solder ball fractional-order parameters; Feature extraction and feature matrix construction are performed on the fractional-order parameters of the solder balls to obtain a multi-scale solder ball feature matrix; Based on the multi-scale solder ball feature matrix, a weighted fusion is performed to obtain the solder ball fusion feature vector; The fused feature vector of the solder ball is input into a pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning.

2. The solder ball quality inspection method based on machine learning according to claim 1, characterized in that, The step of performing fractional wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients specifically includes: Obtain a preset fractional wavelet basis function, and perform a discrete Fourier transform on the standard solder ball waveform data and the preset fractional wavelet basis function to obtain first frequency domain data and second frequency domain data; Perform frequency domain fractional order calculation on the first frequency domain data to obtain frequency domain signal data; The second frequency domain data is subjected to frequency domain fractional order calculation to obtain the basis function data; The multi-scale solder ball wavelet coefficients are obtained by performing inverse discrete Fourier transform on the frequency domain signal data and the basis function data.

3. The solder ball quality inspection method based on machine learning according to claim 1, characterized in that, The step of adaptively selecting the order of the multi-scale solder ball wavelet coefficients to obtain the fractional-order parameters of the solder ball specifically includes: Energy calculations are performed on the wavelet coefficients of the multi-scale solder balls to obtain scale energy information; Obtain the preset fractional order range information, and construct a relational model based on the scale energy information and the fractional order range information to obtain a mapping relational model; Based on the mapping relationship model, the objective function is defined to obtain an effective objective function; The fractional-order parameters of the solder balls are obtained by optimizing the solution based on the effective objective function.

4. The solder ball quality inspection method based on machine learning according to claim 1, characterized in that, The step of extracting features and constructing a feature matrix from the fractional-order parameters of the solder balls to obtain a multi-scale solder ball feature matrix specifically includes: Single-scale feature extraction is performed on the fractional-order parameters of the solder ball to obtain the solder ball scale feature vector; A multi-scale feature set is obtained by constructing a scale feature set based on the solder ball scale feature vector. The feature matrix is ​​constructed based on the multi-scale feature set to obtain the multi-scale solder ball feature matrix.

5. The solder ball quality inspection method based on machine learning according to claim 1, characterized in that, The step of weighted fusion based on the multi-scale solder ball feature matrix to obtain the solder ball fusion feature vector specifically includes: Information entropy will be calculated on the multi-scale solder ball feature matrix to obtain multi-scale information entropy, and weights will be calculated based on the multi-scale information entropy to obtain information entropy weight information. The correlation between the multi-scale solder ball feature matrix and the preset quality level information will be calculated to obtain the correlation coefficient, and the weight will be calculated based on the correlation coefficient to obtain the correlation weight information. Based on the information entropy weight information and the relevance weight information, a comprehensive weight calculation is performed to obtain effective weight information; The solder ball fusion feature vector is obtained by performing weighted fusion calculation based on the multi-scale solder ball feature matrix and the effective weight information.

6. The solder ball quality inspection method based on machine learning according to claim 1, characterized in that, Before the step of inputting the fused feature vector of the solder balls into the pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning, the following steps are also included: Obtain the label data extraction identifier, and extract the label solder ball waveform dataset from the database based on the label data extraction identifier; The labeled solder ball waveform dataset is preprocessed and the dataset is divided into standard training dataset and standard validation dataset. The standard training dataset is input into a pre-built neural network model framework for supervised learning training to obtain an initial quality classification model; The initial quality classification model is adjusted and optimized based on the standard validation dataset until the adjusted and optimized initial quality classification model meets the preset model quality conditions, and the initial quality classification model that meets the preset model quality conditions is taken as the quality classification model.

7. The solder ball quality inspection method based on machine learning according to claim 1, characterized in that, The step of inputting the fused feature vector of the solder ball into a pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning specifically includes: Based on the convolutional layer, convolutional feature extraction is performed on the fused feature vector of the solder ball to obtain the convolutional feature of the solder ball. The convolutional features of the solder balls are expanded by pooling based on the pooling layer to obtain a one-dimensional vector of solder ball features. The one-dimensional vector of solder ball features is transformed by a fully connected layer to obtain the solder ball quality detection result based on machine learning.

8. A solder ball quality inspection device based on machine learning, characterized in that, include: The data processing module is used to acquire solder ball waveform data and perform normalization, filtering, and time window alignment processing on the solder ball waveform data to obtain standard solder ball waveform data. The data transformation module is used to perform fractional wavelet transform on the standard solder ball waveform data to obtain multi-scale solder ball wavelet coefficients. The order selection module is used to adaptively select the order of the multi-scale solder ball wavelet coefficients to obtain the solder ball fractional-order parameters. The matrix construction module is used to extract features and construct feature matrices from the fractional-order parameters of the solder balls to obtain a multi-scale solder ball feature matrix. The weighted fusion module is used to perform weighted fusion based on the multi-scale solder ball feature matrix to obtain a solder ball fusion feature vector. The result acquisition module is used to input the fused feature vector of the solder ball into a pre-trained quality classification model to obtain the solder ball quality detection result based on machine learning.

9. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the machine learning-based solder ball quality inspection method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the machine learning-based solder ball quality inspection method as described in any one of claims 1 to 7.

Images