Acoustic emission online evaluation method and system for bonding strength of copper sheet

Abstract

The invention relates to the technical field of microelectronic packaging, in particular to an acoustic emission online evaluation method and system for copper sheet bonding strength, and the method comprises the steps: arranging a three-dimensional acoustic emission sensing array at a copper sheet bonding station, collecting a multi-dimensional acoustic emission signal in a bonding process, and extracting features of each stage for standardization processing; establishing a bonding strength evaluation model by adopting a transfer learning strategy, determining the bonding quality stability by calculating the bonding quality volatility, and adjusting the bonding pressure according to the bonding quality stability; according to the method, by optimizing the bonding pressure parameter, the purpose of minimizing the standard deviation of the overall bonding strength of the bonding area is achieved, and the stability and reliability of the bonding quality are improved; the system is arranged in a non-contact mode, interference-free monitoring of the bonding process is achieved, the influence of a sensor on the bonding process is avoided, meanwhile, comprehensiveness and accuracy of signal collection are improved, and an effective bonding strength evaluation method is provided in the field of microelectronic packaging.

Description

Technical Field

[0001] The present invention relates to the technical field of microelectronic packaging, in particular to an acoustic emission online evaluation method and system for copper sheet bonding strength. Background Art

[0002] Copper bonding is a common chip connection process in microelectronics packaging manufacturing, and its quality directly impacts product reliability and service life. Traditional bond quality assessment methods rely primarily on destructive testing or offline inspections, such as tensile testing and shear testing. These methods not only result in sample loss but also lack real-time monitoring and timely adjustments, resulting in significant fluctuations in product quality and high defect rates.

[0003] There are some existing non-destructive testing methods, such as ultrasonic scanning and X-ray testing. However, these methods often require specialized equipment, are complex to operate, and are difficult to integrate into production lines for online testing. Furthermore, existing acoustic emission testing technologies often use contact sensors, which can easily interfere with the bonding process and affect measurement accuracy. Furthermore, most techniques only capture acoustic emission signals in a single direction, failing to fully characterize the bond interface.

[0004] Furthermore, existing bond quality assessment models are often based on fixed algorithms or parameter settings, making them difficult to adapt to diverse production environments and material batches, and lacking adaptive learning capabilities. Bond parameter adjustments also often rely on empirical settings, making precise control and closed-loop optimization difficult to achieve.

[0005] Therefore, there is an urgent need to develop a method and system that can realize online evaluation of copper sheet bonding strength and adaptively adjust bonding parameters to improve the consistency and reliability of bonding quality. Summary of the Invention

[0006] The purpose of the present invention is to provide an acoustic emission online assessment method and system for the bond strength of copper sheets. Through the non-contact arrangement of a three-dimensional acoustic emission sensor array, combined with staged multidimensional feature extraction, transfer learning strategy, recursive volatility calculation and standard deviation minimization pressure optimization, accurate online assessment and intelligent control of bond strength can be achieved.

[0007] The present invention proposes an acoustic emission online evaluation method for copper sheet bonding strength, comprising:

[0008] S1: Arrange a three-dimensional acoustic emission sensor array at the copper sheet bonding station, wherein the three-dimensional acoustic emission sensor array includes a first acoustic emission sensor arranged above the copper sheet bonding mold for collecting longitudinal acoustic wave information of the bonding position, a second acoustic emission sensor arranged below the copper sheet bonding mold for collecting transverse acoustic wave information of the bonding position, and an acoustic emission sensor array arranged below the bonding head at a position corresponding to the bonding chip position for collecting vertical acoustic wave information of the bonding position, wherein the acoustic emission sensor array is spaced apart from the copper sheet bonding and does not contact the copper sheet bonding and the copper sheet bonding mold during the bonding process;

[0009] S2: collecting multi-dimensional acoustic emission signals during the bonding process, dividing the bonding process into a heating stage, a pressing stage, and a pressure-applying and holding stage, extracting time domain features, frequency domain features, and time-frequency cross-features of each stage, and performing standardization on the extracted features;

[0010] S3: Using a transfer learning strategy to establish a bond strength assessment model, including: constructing a laboratory dataset using data collected by the bond head pressure sensor and acoustic wave information collected by the first acoustic emission sensor, the second acoustic emission sensor, and the acoustic emission sensor array for model training, and fine-tuning the model on the online dataset to form a bond strength assessment model based on transfer learning;

[0011] S4: calculating a bonding quality fluctuation rate based on the bonding strength evaluation model, determining bonding quality stability according to the fluctuation rate, maintaining current bonding parameters when the fluctuation rate is lower than a first threshold, slightly adjusting bonding pressure when the fluctuation rate is between the first threshold and a second threshold, and significantly adjusting bonding pressure when the fluctuation rate is higher than the second threshold;

[0012] S5: Optimize the bonding pressure parameters through the standard deviation minimization objective function. According to the correlation between the vertical acoustic wave information characteristics of the bonding position and the bond head pressure, calculate the standard deviation of the bond strength at each position in the bonding area, and output the optimal bond head pressure with the goal of minimizing the overall standard deviation of the bond strength in the bonding area.

[0013] Preferably, in step S1, the acoustic emission sensor array is arranged in the gap formed between the bonding copper sheet and the bonding chip, and the acoustic emission sensor array includes an upper acoustic emission sensor array, a middle acoustic emission sensor array and a lower acoustic emission sensor array, which are respectively used to collect acoustic wave information at different depths vertical to the bonding position, thereby improving the integrity and accuracy of the acoustic wave information through multi-level collection.

[0014] Preferably, the features extracted in step S2 include:

[0015] Extract time-domain statistical features such as peak value, mean, variance, median, skewness, kurtosis and standard deviation from the sub-curves of each stage;

[0016] Perform bilateral filtering on the time-amplitude curve to extract frequency distribution features and dominant frequency features;

[0017] Band-pass filtering is performed on the first-order derivative of the time-amplitude curve to extract the shock wave characteristics;

[0018] Among them, a feature selection method combining recursive feature elimination and correlation calculation is adopted to calculate the correlation coefficient between the feature and the bond head pressure. When the correlation coefficient between the feature value and the bond head pressure is greater than the preset threshold, the feature is retained, otherwise it is removed.

[0019] Preferably, the step of constructing the bond strength evaluation model in step S3 includes:

[0020] S31: Collect and establish a laboratory data set of bonding pressure and acoustic emission multi-dimensional acoustic wave information, establish a bonding quality grade, assign grades 1 and 2 to high and low bonding strengths respectively, use the bonding quality grade of the bonded copper sheet as a reference signal, and annotate the collected data with the bonding quality grade;

[0021] S32: Construct a transfer learning neural network model, including a time domain feature processing branch and a frequency domain feature processing branch. The time domain features are processed by a recurrent neural network, and the frequency domain features are processed by a convolutional neural network. The features extracted by the two branches are fused and connected to a fully connected layer. The bond head pressure signal collected by the bond head pressure sensor is used as input, multi-dimensional acoustic wave information is used as output, and the bond quality level is used as a label for training.

[0022] S33: Verify and test the optimized transfer learning neural network model on laboratory datasets and online datasets, and fine-tune the parameters of the transfer learning neural network model on the online dataset to achieve knowledge transfer from the laboratory environment to the production environment.

[0023] Preferably, the bond mass fluctuation rate calculation in step S4 adopts a recursive prediction method, specifically including:

[0024] The bond quality prediction value is calculated based on the current moment and historical data. The current prediction value is determined by the historical bond quality mean, the bond quality value of the corresponding sample at the previous moment, and the quality change of the newly added sample. The influence of historical data and new data is balanced by an adaptive weight coefficient.

[0025] The fluctuation rate was calculated based on the predicted bonding mass values ​​of multiple consecutive samples. When the fluctuation rate was lower than 10%, it was determined to be a stable region; when the fluctuation rate was between 10% and 20%, it was determined to be a fluctuating region; and when the fluctuation rate was higher than 20%, it was determined to be an unstable region.

[0026] As a preferred method, the bonding pressure adjustment method in step S5 is:

[0027] Standardize the collected bond head pressure and vertical acoustic wave information of the bonding position;

[0028] Calculate the Pearson correlation coefficient between the normalized vertical acoustic wave information characteristics and the bond head pressure;

[0029] Sort the acoustic wave information features vertical to the bonding position in descending order of correlation coefficient, and take the first n acoustic wave information feature values ​​to construct a subset;

[0030] The relationship between the bond head pressure and each eigenvalue corresponding to the subset is calculated using the least squares method;

[0031] Taking the minimum standard deviation of the overall bond strength in the bonding area as the objective function, the optimal bond head pressure is output;

[0032] When the bonding quality is in the fluctuating zone, adjust the bonding pressure by 5% to 7%; when the bonding quality is in the unstable zone, adjust the bonding pressure by 8% to 10% until the bonding quality reaches the stable zone.

[0033] Preferably, the first acoustic emission sensor, the second acoustic emission sensor and the acoustic emission sensing array use different types of sensors according to the collection position and signal characteristics, including piezoelectric ceramic sensors for high-frequency signal collection, piezoelectric film sensors for wide-band signal collection and piezoelectric cantilever beam sensors for low-frequency signal collection.

[0034] Preferably, 3 to 12 first acoustic emission sensors, 3 to 12 second acoustic emission sensors, and 30 to 120 acoustic emission sensor arrays are arranged, and the high-density sensor array is used to improve the spatial resolution and data redundancy of acoustic emission signal acquisition.

[0035] As an example, after step S5 is completed, the following steps are further included:

[0036] outputting the adjusted bond head pressure corresponding to the pressure to the bonding device;

[0037] After performing the bonding process, verify the bond quality;

[0038] Feedback the verification results to the bond strength evaluation model to achieve continuous optimization of the model;

[0039] Establish a historical database of bonding parameters and bonding quality for continuous improvement of subsequent bonding processes.

[0040] Acoustic emission online assessment system for copper sheet bond strength, including:

[0041] The data acquisition module includes a first acoustic emission sensor arranged above the bonding copper sheet mold, a second acoustic emission sensor arranged below the bonding copper sheet mold, an acoustic emission sensor array arranged below the bonding head at a position corresponding to the bonding chip position, and a bonding head pressure sensor. The first acoustic emission sensor and the second acoustic emission sensor are respectively used to collect longitudinal and transverse acoustic wave information of the bonding position. The acoustic emission sensor array is used to collect vertical acoustic wave information of the bonding position. The acoustic emission sensor array is spaced apart from the bonding copper sheet and does not contact the bonding copper sheet or the bonding copper sheet mold during the bonding process.

[0042] Feature processing module: used to divide the bonding process into heating stage, pressing stage and pressure holding stage, respectively extract the time domain features, frequency domain features and time-frequency cross features of each stage, standardize the extracted features, and select the optimal feature subset by combining recursive feature elimination method with correlation calculation;

[0043] Model building module: This module uses a transfer learning strategy to construct a laboratory dataset using data collected by the bond head pressure sensor and multi-dimensional acoustic wave information for model training. It then performs fine-tuning on the online dataset to form a bond strength assessment model.

[0044] Quality assessment module: used to calculate the bond quality fluctuation rate based on the bond strength assessment model, determine the bond quality stability based on the fluctuation rate, and judge whether the bond quality is in the stable zone, fluctuation zone or unstable zone;

[0045] Pressure adjustment module: used to optimize the bonding pressure parameters through the standard deviation minimization objective function. According to the correlation between the vertical acoustic wave information characteristics of the bonding position and the bond head pressure, the standard deviation of the bond strength at each position in the bonding area is calculated. The optimal bond head pressure is output with the goal of minimizing the overall standard deviation of the bond strength in the bonding area. The corresponding pressure adjustment strategy is executed according to the area where the bonding quality is located.

[0046] The acoustic emission online evaluation method and system for copper sheet bonding strength provided by the present invention have the following beneficial effects:

[0047] 1. The non-contact arrangement of the three-dimensional acoustic emission sensor array enables interference-free monitoring of the bonding process, avoiding the impact of the sensor on the bonding process while improving the comprehensiveness and accuracy of signal acquisition;

[0048] 2. A phased multidimensional feature extraction method is used to extract time domain, frequency domain, and time-frequency cross-features for different stages of the bonding process. Recursive feature elimination and correlation calculation are combined to screen out the most representative features, improving the accuracy and robustness of the evaluation model.

[0049] 3. Introducing a transfer learning strategy to transfer knowledge trained in the laboratory environment to the production environment, solving the problem of insufficient data in the early stages of production while improving the model's adaptability to environmental and material changes;

[0050] 4. Through the recursive volatility calculation method, dynamic prediction and fluctuation monitoring of bonding quality are achieved, which is highly sensitive to quality fluctuations and can detect potential problems in a timely manner;

[0051] 5. Based on the pressure optimization strategy of minimizing the standard deviation, while considering the uniformity and average strength of the bonding strength, precise regulation of the bonding parameters is achieved, forming a closed-loop control of quality assessment and parameter adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flow chart of the method for online acoustic emission evaluation of copper sheet bond strength provided by an embodiment of the present invention;

[0053] Figure 2 A flowchart of multi-dimensional feature extraction in stages provided by an embodiment of the present invention;

[0054] Figure 3 A structural diagram of a transfer learning neural network model provided by an embodiment of the present invention;

[0055] Figure 4 A schematic diagram of recursive volatility calculation and quality prediction provided by an embodiment of the present invention;

[0056] Figure 5 A schematic diagram of a standard deviation minimization pressure optimization process provided by an embodiment of the present invention;

[0057] Figure 6 This is a structural block diagram of the copper sheet bond strength online acoustic emission evaluation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] Please refer to Figure 1 - Figure 6 The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0059] like Figure 1 As shown, the method includes the following steps:

[0060] Step S1, arranging a three-dimensional acoustic emission sensor array:

[0061] A three-dimensional acoustic emission sensor array is arranged at the copper sheet bonding station, and the three-dimensional acoustic emission sensor array includes a first acoustic emission sensor arranged above the copper sheet bonding mold, a second acoustic emission sensor arranged below the copper sheet bonding mold, and an acoustic emission sensor array arranged below the bonding head at a position corresponding to the bonding chip position.

[0062] The first acoustic emission sensor is arranged above the bonding copper sheet mold to collect longitudinal acoustic wave information of the bonding position; the second acoustic emission sensor is arranged below the bonding copper sheet mold to collect transverse acoustic wave information of the bonding position; the acoustic emission sensor array is arranged below the bonding head corresponding to the bonding chip position to collect vertical acoustic wave information of the bonding position.

[0063] A key innovation of the present invention is that the acoustic emission sensor array maintains a certain distance from the copper bonding sheet, and does not contact the copper bonding sheet or the bonding mold during the bonding process. This non-contact measurement method avoids sensor interference with the bonding process, while improving signal quality and measurement accuracy. In a preferred embodiment of the present invention, this distance is maintained between 0.05 mm and 0.5 mm, a range that ensures that the sensor does not contact the bonding part while ensuring effective transmission of the acoustic wave signal.

[0064] In one embodiment of the present invention, the acoustic emission sensor array is arranged in the gap formed between the bonding copper sheet and the bonding chip, and is further subdivided into an upper acoustic emission sensor array, a middle acoustic emission sensor array and a lower acoustic emission sensor array, which are respectively used to collect acoustic wave information at different depths vertical to the bonding position, thereby improving the integrity and accuracy of the acoustic wave information through multi-level collection.

[0065] Preferably, 3 to 12 first acoustic emission sensors, 3 to 12 second acoustic emission sensors, and 30 to 120 acoustic emission sensor arrays are arranged. This high-density sensor array improves the spatial resolution and data redundancy of acoustic emission signal acquisition. Specifically, 6 first acoustic emission sensors can be arranged in a regular hexagonal pattern, 6 second acoustic emission sensors can be arranged in a pattern corresponding to the first acoustic emission sensors, and 64 acoustic emission sensor arrays can be arranged in an 8×8 matrix pattern. This arrangement achieves comprehensive coverage and uniform sampling of the bonding area.

[0066] The first and second acoustic emission sensors, as well as the acoustic emission sensor array described herein, utilize different types of sensors based on the acquisition location and signal characteristics. These include piezoelectric ceramic sensors for high-frequency signal acquisition, piezoelectric film sensors for broadband signal acquisition, and piezoelectric cantilever beam sensors for low-frequency signal acquisition. In practical applications, the first and second acoustic emission sensors may utilize piezoelectric ceramic sensors with a frequency response range of 50kHz-500kHz; the upper and lower acoustic emission sensor arrays may utilize piezoelectric film sensors with a frequency response range of 10kHz-2MHz; and the middle acoustic emission sensor array may utilize piezoelectric cantilever beam sensors with a frequency response range of 1kHz-100kHz. This heterogeneous sensor combination can capture acoustic emission signals in different frequency bands, providing more comprehensive information about the bonding interface.

[0067] Step S2, multi-dimensional acoustic emission signal acquisition and feature extraction:

[0068] Multi-dimensional acoustic emission signals are collected during the bonding process. The bonding process is divided into a heating stage, a pressing stage, and a pressure-holding stage. Time domain features, frequency domain features, and time-frequency cross-features of each stage are extracted, and the extracted features are standardized.

[0069] like Figure 2 As shown, the bonding process is first precisely divided into three stages: heating (T1-T2), pressing (T2-T3), and pressurization and holding (T3-T4). This division is based on the varying characteristics of bond head pressure and temperature. Typically, the heating stage is characterized by a rapid temperature increase and low pressure; the pressing stage is characterized by a rapid pressure increase; and the pressurization and holding stages are characterized by relatively stable pressure and temperature. In the actual copper sheet bonding process, the temperature during the heating stage typically rises from room temperature to 200-300°C for 1-3 seconds; the pressure during the pressing stage increases from 0.1 MPa to 1-5 MPa for 0.2-1 second; and the pressure and temperature during the pressurization and holding stages remain constant for 1-5 seconds.

[0070] For each stage, the present invention extracts three types of features:

[0071] Time domain features: Extract statistical features such as peak value, mean value, variance, median value, skewness, kurtosis and standard deviation from each stage sub-curve. In this embodiment, the time domain feature extraction formula is as follows:

[0072] ,

[0073] in: is the mean value of the acoustic emission signal, in volts (V); The acoustic emission signal The amplitude of each sampling point, in volts (V); is the total number of sampling points, usually 1000 to 10000 points.

[0074] variance:

[0075] ,

[0076] in: is the variance of the acoustic emission signal, in volts squared (V^{2}); The acoustic emission signal The amplitude of each sampling point, in volts (V); is the signal mean value, in volts (V); is the total number of sampling points.

[0077] Kurtosis:

[0078] ,

[0079] in: is the kurtosis of the acoustic emission signal, dimensionless; The acoustic emission signal The amplitude of each sampling point, in volts (V); is the signal mean value, in volts (V); is the signal standard deviation, in volts (V); is the total number of sampling points. Kurtosis describes the sharpness of the signal distribution. For a normal distribution, the kurtosis value is 3.

[0080] Skewness:

[0081] ,

[0082] in: is the skewness of the acoustic emission signal, dimensionless; The acoustic emission signal The amplitude of each sampling point, in volts (V); is the signal mean value, in volts (V); is the signal standard deviation, in volts (V); is the total number of sampling points. Skewness describes the symmetry of the signal distribution, with positive values ​​indicating right skewness and negative values ​​indicating left skewness.

[0083] In the copper sheet bonding process, these time domain features are strongly correlated with the bonding quality. For example, high-quality bonding is usually characterized by a low peak value (<0.5V) of the acoustic emission signal and a uniform distribution with a small variance (<0.05V). 2） , with a kurtosis close to 3; low-quality bonding is usually characterized by a high peak (>1V) and uneven acoustic emission signal with a large variance (>0.2V 2） , the kurtosis is much greater than 3.

[0084] Frequency domain features: Perform bilateral filtering on the time-amplitude curve to extract frequency distribution features and dominant frequency features. Frequency domain feature extraction first converts the time domain signal into the frequency domain through fast Fourier transform (FFT):

[0085] ,

[0086] in: The frequency domain signal The complex value of the frequency points; The time domain signal The amplitude of each sampling point, in volts (V); is the number of FFT points, usually 4096 or 8192; is an imaginary unit; is the base of natural logarithms; is pi; and are the indices of frequency domain and time domain respectively.

[0087] Then calculate the power spectral density:

[0088] ,

[0089] in: For the The power spectral density of each frequency point, in watts per hertz (W / Hz); is the complex value of the kth frequency point of the frequency domain signal; Indicates plural The modulus of ; N is the number of FFT points.

[0090] Main frequency feature extraction:

[0091] ,

[0092] in: is the main frequency of the acoustic emission signal, in Hertz (Hz); Indicates Frequency domain index with the maximum value ; is the sampling frequency, in Hertz (Hz), and 2 MHz is used in this embodiment; is the number of FFT points.

[0093] Band energy ratio:

[0094] ,

[0095] in: It is the ratio of the energy in a specific frequency band to the total energy, dimensionless; For the The power spectral density of each frequency point; and Respectively represent the start and end indexes of a specific frequency band; In copper bonding applications, the 50-200 kHz frequency band is often of interest, as it contains key information about the formation of the bonding interface.

[0096] During copper bonding, different bond qualities exhibit distinct frequency domain characteristics. High-quality bonds typically exhibit a dominant frequency in the 100-150kHz range and an energy ratio between 0.6 and 0.8 in the 50-200kHz band. Low-quality bonds, on the other hand, exhibit unstable dominant frequencies, with energy ratios below 0.5 or scattered distribution in the 50-200kHz band.

[0097] Time-frequency cross-feature: Bandpass filter the first-order derivative of the time-amplitude curve to extract the shock wave characteristics. The time-frequency cross-feature first calculates the first-order derivative of the time domain signal:

[0098] ,

[0099] in: is a time domain signal The first derivative of , in volts per second (V / s); and They are the time domain signals and The amplitude of each sampling point, in volts (V).

[0100] The derivative signal is then bandpass filtered, typically within a 10kHz-500kHz filter band. This frequency range contains critical shock wave information during the bonding process. In copper bonding, shock wave characteristics are closely related to the formation of the bond interface. High-quality bonds typically exhibit a significant shock wave signal spike (amplitude >0.2V / s) early in the bonding phase, followed by rapid decay. Low-quality bonds, on the other hand, typically exhibit a low shock wave signal amplitude (<0.1V / s) or an irregular distribution.

[0101] For all extracted features, the present invention adopts a feature selection method that combines recursive feature elimination with correlation calculation to calculate the correlation coefficient between the feature and the bond head pressure. When the correlation coefficient between the feature value and the bond head pressure is greater than a preset threshold, the feature is retained; otherwise, it is removed. The correlation coefficient is calculated using the Pearson correlation coefficient:

[0102] ,

[0103] in: Characterized by Bond head pressure The Pearson correlation coefficient between them is dimensionless and ranges from [-1,1]; For the The characteristic value of each sample; For the The bond head pressure value of each sample, in MPa; and represent the characteristic value and the average value of the bond head pressure, respectively; In practical applications, the correlation threshold is usually set to 0.6-0.8, and is preferably set to 0.7 in the present invention. This value can achieve a good balance between retaining key features and removing redundant features.

[0104] The selected features are standardized using the Z-score standardization method:

[0105] ,

[0106] in: is the normalized eigenvalue, dimensionless; is the original eigenvalue; is the characteristic mean; is the feature standard deviation. Standardization can eliminate the influence of different feature dimensions and improve the stability and accuracy of model training.

[0107] Step S3, transfer learning strategy and bond strength evaluation model construction:

[0108] A transfer learning strategy is adopted to establish a bond strength evaluation model, including: constructing a laboratory dataset for model training by combining the data collected by the bond head pressure sensor with the acoustic wave information collected by the first acoustic emission sensor, the second acoustic emission sensor and the acoustic emission sensor array, and fine-tuning it on the online dataset to form a bond strength evaluation model based on transfer learning.

[0109] like Figure 3 As shown, the model includes the following construction steps:

[0110] First, a laboratory dataset of bonding pressure and acoustic emission multidimensional acoustic wave information was collected and established. Bond quality grades were then established, with high and low bond strengths assigned grades of 1 and 2, respectively. The bond quality grade of the copper sheet was used as a reference signal, and the collected data was labeled with the bond quality grade. In practice, bond strength standards are determined through tensile or shear tests. Typically, bond strengths greater than 10 MPa are classified as Grade 1 (high strength), and bond strengths less than 5 MPa are classified as Grade 2 (low strength). In copper sheet bonding applications, joints with bond strengths greater than 10 MPa demonstrate good reliability in thermal cycling and vibration testing, while joints with strengths less than 5 MPa are prone to failure under environmental stress.

[0111] Secondly, a transfer learning neural network model is constructed, including a time domain feature processing branch and a frequency domain feature processing branch. The time domain features are processed by a recurrent neural network (LSTM), and the frequency domain features are processed by a convolutional neural network (CNN). The features extracted by the two branches are fused and connected to the fully connected layer. The bond head pressure signal collected by the bond head pressure sensor is used as input, multi-dimensional acoustic wave information is used as output, and the bonding quality grade is used as the label for training.

[0112] The time domain feature processing branch uses the LSTM network, which is mathematically expressed as follows:

[0113] Input Gate:

[0114] ,

[0115] in: is the state of the input gate at time t, with a value range of [0,1], which controls the proportion of new input information entering the memory unit; The sigmoid activation function maps the input to the [0,1] interval; is the weight matrix of the input gate, with dimension ,in is the hidden state dimension, is the input dimension; is the hidden state at time t-1, with dimension ; is the input at time t, and its dimension is ; Indicates that and Splice by column; is the bias vector of the input gate, with dimension ; Represents matrix multiplication.

[0116] Forget Gate:

[0117] ,

[0118] in: is the forget gate state at time t, with a value range of [0,1], which controls the proportion of historical information retained; is the sigmoid activation function; is the weight matrix of the forget gate, with dimension ; is the hidden state at time t-1; is the input at time t; is the bias vector of the forget gate, with dimension .

[0119] Output Gate:

[0120] ,

[0121] in: is the output gate state at time t, with a value range of [0,1], which controls the ratio of information output of the memory unit; is the sigmoid activation function; is the weight matrix of the output gate, with dimension ; is the hidden state at time t-1; is the input at time t; is the bias vector of the output gate, with dimension .

[0122] Candidate memory cells:

[0123] ,

[0124] in: is the candidate memory unit state at time t, with a value range of [-1,1]; It is a hyperbolic tangent activation function that maps the input to the interval [-1,1]; is the weight matrix of the candidate memory unit, with dimension ; is the hidden state at time t-1; is the input at time t; is the bias vector of the candidate memory unit, with dimension .

[0125] Memory unit update:

[0126] ,

[0127] in: is the state of the memory unit at time t; is the forget gate state at time t; is the state of the memory unit at time t-1; is the input gate state at time t; is the candidate memory cell state at time t; * represents element-by-element multiplication (Hadamard product).

[0128] Hidden state output:

[0129] ,

[0130] in: is the hidden state output at time t, with dimension ; is the output gate state at time t; is the state of the memory unit at time t; is the hyperbolic tangent activation function; * indicates element-wise multiplication.

[0131] Hidden state dimension of LSTM network in copper bonding application Usually set to 64 or 128, input dimension The number of time-domain features is determined, typically 10 to 20. LSTM networks can effectively capture temporal dependencies in the bonding process, and are particularly good at identifying feature changes at different bonding stages.

[0132] The frequency domain feature processing branch uses the CNN network, and its mathematical expression is as follows:

[0133] Convolution operation:

[0134] ,

[0135] in: Indicates the Tier Linear output of feature maps; Indicates the The number of feature maps of the layer; Indicates the Layer from the Layer feature maps to the The convolution kernel of the feature map; Indicates the Layer feature maps; Indicates the corresponding bias; represents a two-dimensional convolution operation, defined as:

[0136] ,

[0137] in: and Represent the height and width of the convolution kernel respectively; Indicates that the convolution kernel is at position The weight of Indicates that the feature map is at position The value of .

[0138] Activation function:

[0139] ,

[0140] in: Represents the activated feature map; ReLU represents the rectified linear unit activation function, defined as ReLU .

[0141] Pooling operation:

[0142] ,

[0143] in: Represents the feature map after pooling; MaxPool represents the maximum pooling operation, which is defined as:

[0144] ,

[0145] in: and Represent the height and width of the pooling window respectively; and They represent the pooling stride respectively.

[0146] In copper bonding applications, CNN networks typically contain three to four convolutional layers, with the number of convolution kernels in each layer ranging from 16 to 64, kernel sizes of 3×3 or 5×5, and a pooling window size of 2×2. CNN networks are effective in extracting local patterns and spatial structures from frequency domain features, and are particularly effective in identifying frequency features during the formation of the bonding interface.

[0147] The features of the two branches are fused using a concatenation operation and then classified through a fully connected layer:

[0148] Feature fusion:

[0149] ,

[0150] in: Represents the fused feature vector, with dimension ; Represents the output of the last time step of LSTM, with dimension ; Represents the features after the last layer of CNN pooling, the dimension is ; Represents a vector concatenation operation.

[0151] Fully connected layer:

[0152] ,

[0153] in: Represents the output of the fully connected layer, with dimension , is the number of categories (2 in this example); Represents the weight matrix of the fully connected layer, with dimension ; Represents the fused feature vector; Represents the bias vector, dimension is .

[0154] Output layer:

[0155] ,

[0156] in: Represents the probability distribution of bonding quality grade predicted by the model, with dimension ; represents the softmax activation function, which is defined as:

[0157] ,

[0158] in: Indicates the Scores for each category; is the base of natural logarithms.

[0159] Finally, the optimized transfer learning neural network model is validated and tested on both laboratory and online datasets. The parameters of the transfer learning neural network model are fine-tuned on the online dataset, enabling the transfer of knowledge from the laboratory environment to the production environment. The core idea of ​​transfer learning is to retain the weight parameters of the low- and mid-layers of the network, which capture general feature representations, while only fine-tuning the parameters of the high-layer network to adapt to the specific characteristics of the new environment.

[0160] In copper bonding applications, laboratory datasets typically contain 500 to 1,000 samples. These samples are collected under strictly controlled conditions and cover different combinations of bonding parameters (pressure, temperature, and time). Online datasets, on the other hand, are typically collected from actual production lines and may initially contain only 50 to 100 samples. Through transfer learning, the model can quickly adapt to the actual production environment using limited online data, reducing the need for large amounts of labeled data.

[0161] During model training, the cross-entropy loss function and the Adam optimizer were used. The initial learning rate was set to 0.001, the batch size was set to 64, and the number of training epochs was 100. To prevent overfitting, an early stopping strategy and L2 regularization were employed, with the regularization coefficient set to 0.0001. These parameter settings were determined based on extensive experiments and achieve a good balance between training efficiency and model performance.

[0162] Step S4, recursive volatility calculation and quality stability assessment:

[0163] The bonding quality fluctuation rate is calculated based on the bonding strength evaluation model, and the bonding quality stability is determined according to the fluctuation rate. When the fluctuation rate is lower than a first threshold, the current bonding parameters are maintained. When the fluctuation rate is between the first threshold and the second threshold, the bonding pressure is slightly adjusted. When the fluctuation rate is higher than the second threshold, the bonding pressure is greatly adjusted.

[0164] Figure 5 FIG. 1 shows a schematic diagram of recursive volatility calculation and quality prediction in an embodiment of the present invention. Figure 5 As shown, the bond quality fluctuation rate is calculated using a recursive prediction method to calculate the bond quality prediction value based on the current time and historical data:

[0165] ,

[0166] in: represents the predicted bonding mass corresponding to the i-th sample at time t, in megapascals (MPa); It represents the average bonding mass of n samples before time t-1, in MPa, and is calculated as represents the bonding mass corresponding to the i-th sample at time t-1, in megapascals (MPa); Indicates the mass change of the newly added sample at time t relative to time t-1, in megapascals (MPa); Represents the adaptive weight coefficient, dimensionless, ranging from 0.1 to 0.5, dynamically adjusted according to the current production stability, and the calculation formula is ,in is a tuning parameter, usually set to 1.0, is the volatility at time t-1.

[0167] In copper sheet bonding applications, n is usually set to 5-10, which means considering the average of the most recent 5 to 10 samples; It can be calculated by the relationship between the current bonding parameters and the model prediction, usually within the range of ±1MPa; adaptive weight The design makes it possible to consider the group mean more when the fluctuation is large (to improve stability) and consider the individual historical value more when the fluctuation is small (to maintain continuity).

[0168] Calculate the volatility based on the bond mass predictions for several consecutive samples:

[0169] ,

[0170] in: Indicates volatility in percentage (%); It represents the standard deviation of the bonding quality in MPa. The calculation formula is: ,in The number of samples to be considered is usually 10-20; It represents the average value of the bonding quality in megapascals (MPa), and the calculation formula is .

[0171] According to the fluctuation rate, the bonding quality stability is divided into three areas:

[0172] When the volatility is lower than 10%, it is determined to be in a stable area and the current bonding parameters are kept unchanged;

[0173] When the fluctuation rate is between 10% and 20%, it is determined to be in the fluctuation zone and the bonding pressure needs to be slightly adjusted;

[0174] When the fluctuation rate is higher than 20%, it is judged to be in an unstable area and the bonding pressure needs to be significantly adjusted.

[0175] These three thresholds (10% and 20%) were determined through extensive production practice and reflect the actual conditions of the copper sheet bonding process, effectively balancing control sensitivity and system stability. In copper sheet bonding production, a fluctuation rate below 10% generally indicates stable and consistent bond quality; a fluctuation rate between 10% and 20% indicates some fluctuation in bond quality, but remains within an acceptable range; and a fluctuation rate above 20% indicates significant fluctuation in bond quality, potentially indicating material or process issues and requiring timely adjustments.

[0176] Step S5, standard deviation minimization pressure optimization and parameter adjustment:

[0177] The bonding pressure parameters are optimized through the standard deviation minimization objective function. According to the correlation between the vertical acoustic wave information characteristics of the bonding position and the bond head pressure, the standard deviation of the bond strength at each position in the bonding area is calculated. The optimal bond head pressure is output with the goal of minimizing the standard deviation of the overall bond strength in the bonding area.

[0178] like Figure 5As shown, the collected bond head pressure and the vertical acoustic wave information of the bonding position are first normalized, and the Pearson correlation coefficient between the normalized vertical acoustic wave information characteristics and the bond head pressure is calculated:

[0179] ,

[0180] in: Features Bond head pressure The Pearson correlation coefficient between them is dimensionless and ranges from [-1,1]; For the The standardized acoustic information characteristics of the samples; For the The normalized bond head pressure of each sample, in megapascals (MPa); and They represent the corresponding average values ​​respectively; Indicates the sample size.

[0181] The acoustic wave information features perpendicular to the bonding position are sorted in descending order by correlation coefficient, and the top n acoustic wave information feature values ​​are selected to form a subset. The value of n is usually 20% to 30% of the total number of features, and in this embodiment, 25% is preferred. This ratio can retain most of the key information while effectively reducing computational complexity. For example, if the total number of original features is 40, the top 10 features with the highest correlation are selected to form a subset.

[0182] The relationship between the bond head pressure and each eigenvalue corresponding to the subset is calculated using the least squares method:

[0183] ,

[0184] in: represents the predicted bond head pressure in megapascals (MPa); Indicates the eigenvalues; Indicates the corresponding weight; Represents the bias term. Weight and bias Solved by minimizing the mean square error:

[0185] ,

[0186] in: Indicates the actual bond head pressure in megapascals (MPa); It represents the predicted bond head pressure in megapascals (MPa); represents the number of samples. The solution to the least squares problem can be solved by the normal equation:

[0187] ,

[0188] in: is a vector containing all weights and biases, with dimension ; is the feature matrix with dimension , where the first column is all 1 (corresponding to the bias term), and the following Listed as eigenvalues; is the actual bond head pressure vector, with dimension ; express The transpose of express The inverse matrix of .

[0189] The optimal bond head pressure is determined by minimizing the standard deviation of the overall bond strength in the bonding area as the objective function:

[0190] ,

[0191] in: Indicates the optimal bond head pressure in megapascals (MPa); Expressing pressure Next The bond strength at each position, in megapascals (MPa); Expressing pressure The average bond strength under the condition of , in megapascals (MPa), is calculated as ; Indicates the number of measurement locations in the bonding area.

[0192] In actual copper sheet bonding applications, It can be predicted from the acoustic wave information characteristics through the mapping relationship established above. The pressure range usually explored is 1-5MPa, with a step size of 0.1MPa, that is, the pressure value that minimizes the standard deviation of the bonding strength is found within this range.

[0193] Implement corresponding pressure adjustment strategy according to the bond quality stability area:

[0194] When the bonding quality is in the fluctuation zone, adjust the bonding pressure by 5% to 7%;

[0195] When the bonding quality is in the unstable zone, adjust the bonding pressure by 8% to 10%;

[0196] Continue adjusting until the bonding quality reaches a stable zone.

[0197] In practical applications, the adjustment range for the fluctuation zone is preferably 6%, and the adjustment range for the unstable zone is preferably 9%. These values ​​are determined based on extensive experimental data and production experience, and can achieve a good balance between response speed and system stability. For example, if the current bonding pressure is 3MPa and the bond quality is in the fluctuation zone, the adjustment range is 3MPa × 6% = 0.18MPa, which can adjust the pressure to 3.18MPa or 2.82MPa. The specific direction is determined by the aforementioned standard deviation minimization objective function.

[0198] After step S5 is completed, it further includes: outputting the adjusted bonding head pressure corresponding pressure to the bonding equipment; verifying the bonding quality after executing the bonding process; feeding back the verification results to the bonding strength evaluation model to achieve continuous optimization of the model; establishing a historical database of bonding parameters and bonding quality for continuous improvement of subsequent bonding processes.

[0199] like Figure 6 As shown, the system includes a data acquisition module 100 , a feature processing module 200 , a model building module 300 , a quality assessment module 400 and a pressure adjustment module 500 .

[0200] The data acquisition module 100 includes a first acoustic emission sensor 110 arranged above the bonding copper sheet mold, a second acoustic emission sensor 120 arranged below the bonding copper sheet mold, an acoustic emission sensor array 130 arranged below the bonding head and corresponding to the bonding chip position, and a bonding head pressure sensor 140.

[0201] The first acoustic emission sensor 110 and the second acoustic emission sensor 120 are respectively used to collect longitudinal and transverse acoustic wave information of the bonding position, and the acoustic emission sensor array 130 is used to collect vertical acoustic wave information of the bonding position. The acoustic emission sensor array 130 maintains a gap between itself and the bonding copper sheet and does not contact the bonding copper sheet or the bonding copper sheet mold during the bonding process.

[0202] In an embodiment of the present invention, six first acoustic emission sensors 110 and six second acoustic emission sensors 120 are arranged in a regular hexagonal distribution, 64 acoustic emission sensor arrays 130 are arranged in an 8×8 matrix distribution, the sampling rate is set to 2MHz, and the sampling accuracy is 16bit. These parameter settings can meet the high-precision acquisition requirements of acoustic emission signals in the bonding process.

[0203] The feature processing module 200 is used to divide the bonding process into a heating stage, a pressing stage, and a pressurization and holding stage, respectively extract the time domain features, frequency domain features, and time-frequency cross-features of each stage, standardize the extracted features, and select the optimal feature subset by combining recursive feature elimination with correlation calculation.

[0204] The feature processing module 200 includes a stage division unit 210, a feature extraction unit 220, and a feature selection unit 230. The stage division unit 210 divides the bonding process into three stages based on the bond head pressure and temperature variation characteristics; the feature extraction unit 220 extracts multidimensional features for each stage; and the feature selection unit 230 selects the optimal feature subset by combining recursive feature elimination with correlation calculation.

[0205] The model building module 300 is used to adopt a transfer learning strategy to construct a laboratory data set based on the data collected by the bond head pressure sensor and the multi-dimensional acoustic wave information for model training, and to perform fine-tuning on the online data set to form a bond strength evaluation model.

[0206] The model construction module 300 includes a dataset construction unit 310, a network construction unit 320, and a transfer learning unit 330. The dataset construction unit 310 is responsible for constructing the laboratory dataset and the online dataset; the network construction unit 320 is responsible for constructing a neural network model that includes a time domain feature processing branch and a frequency domain feature processing branch; and the transfer learning unit 330 is responsible for training the base model on the laboratory dataset and fine-tuning it on the online dataset.

[0207] The quality assessment module 400 is used to calculate the bonding quality fluctuation rate based on the bonding strength assessment model, determine the bonding quality stability according to the fluctuation rate, and judge whether the bonding quality is in a stable zone, a fluctuating zone, or an unstable zone.

[0208] The quality assessment module 400 includes a recursive prediction unit 410 and a stability assessment unit 420. The recursive prediction unit 410 uses a recursive method to predict the bonding quality and calculate the fluctuation rate; the stability assessment unit 420 determines whether the bonding quality is in a stable zone, a fluctuating zone, or an unstable zone based on the fluctuation rate.

[0209] The pressure adjustment module 500 is used to optimize the bonding pressure parameters through the standard deviation minimization objective function. According to the correlation between the vertical acoustic wave information characteristics of the bonding position and the bond head pressure, the standard deviation of the bonding strength at each position in the bonding area is calculated, and the optimal bond head pressure is output with the goal of minimizing the overall standard deviation of the bonding strength in the bonding area. The corresponding pressure adjustment strategy is executed according to the area where the bonding quality is located.

[0210] The pressure adjustment module 500 includes a correlation analysis unit 510, a standard deviation calculation unit 520, and a pressure optimization unit 530. The correlation analysis unit 510 analyzes the correlation between the acoustic wave information characteristics and the bond head pressure; the standard deviation calculation unit 520 calculates the standard deviation of the strength of the bonding area under different pressures; and the pressure optimization unit 530 determines the optimal pressure that minimizes the standard deviation and implements the corresponding adjustment strategy.

[0211] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made by using the contents of the present description and drawings under the inventive concept of the present invention, or direct / indirect application in other related technical fields are included in the scope of patent protection of the present invention.

Claims

1. The acoustic emission online evaluation method for copper sheet bonding strength is characterized by: The following steps are involved: S1: Arrange a three-dimensional acoustic emission sensor array at the copper sheet bonding station, wherein the three-dimensional acoustic emission sensor array includes a first acoustic emission sensor arranged above the copper sheet bonding mold for collecting longitudinal acoustic wave information of the bonding position, a second acoustic emission sensor arranged below the copper sheet bonding mold for collecting transverse acoustic wave information of the bonding position, and an acoustic emission sensor array arranged below the bonding head at a position corresponding to the bonding chip position for collecting vertical acoustic wave information of the bonding position, wherein the acoustic emission sensor array is spaced apart from the copper sheet bonding and does not contact the copper sheet bonding and the copper sheet bonding mold during the bonding process; S2: collecting multi-dimensional acoustic emission signals during the bonding process, dividing the bonding process into a heating stage, a pressing stage, and a pressure-applying and holding stage, extracting time domain features, frequency domain features, and time-frequency cross-features of each stage, and performing standardization on the extracted features; S3: Using a transfer learning strategy to establish a bond strength assessment model, including: constructing a laboratory dataset using data collected by the bond head pressure sensor and acoustic wave information collected by the first acoustic emission sensor, the second acoustic emission sensor, and the acoustic emission sensor array for model training, and fine-tuning the model on the online dataset to form a bond strength assessment model based on transfer learning; S4: calculating a bonding quality fluctuation rate based on the bonding strength evaluation model, determining bonding quality stability according to the fluctuation rate, maintaining current bonding parameters when the fluctuation rate is lower than a first threshold, slightly adjusting bonding pressure when the fluctuation rate is between the first threshold and a second threshold, and significantly adjusting bonding pressure when the fluctuation rate is higher than the second threshold; S5: Optimize the bonding pressure parameters through the standard deviation minimization objective function. According to the correlation between the vertical acoustic wave information characteristics of the bonding position and the bond head pressure, calculate the standard deviation of the bond strength at each position in the bonding area, and output the optimal bond head pressure with the goal of minimizing the overall standard deviation of the bond strength in the bonding area.

2. The acoustic emission online evaluation method for copper sheet bonding strength according to claim 1, characterized in that: In step S1, the acoustic emission sensor array is arranged in the gap formed between the bonding copper sheet and the bonding chip. The acoustic emission sensor array includes an upper acoustic emission sensor array, a middle acoustic emission sensor array and a lower acoustic emission sensor array, which are respectively used to collect acoustic wave information at different depths vertical to the bonding position, thereby improving the integrity and accuracy of the acoustic wave information through multi-level collection.

3. The acoustic emission online evaluation method for copper sheet bonding strength according to claim 1, characterized in that: The features extracted in step S2 include: Extract the time domain statistical features of peak value, mean, variance, median, skewness, kurtosis and standard deviation for the sub-curves of each stage; Perform bilateral filtering on the time-amplitude curve to extract frequency distribution features and dominant frequency features; Band-pass filtering is performed on the first-order derivative of the time-amplitude curve to extract the shock wave characteristics; Among them, a feature selection method combining recursive feature elimination and correlation calculation is adopted to calculate the correlation coefficient between the feature and the bond head pressure. When the correlation coefficient between the feature value and the bond head pressure is greater than the preset threshold, the feature is retained, otherwise it is removed.

4. The acoustic emission online evaluation method for copper sheet bonding strength according to claim 1, characterized in that: The steps of constructing the bond strength evaluation model in step S3 include: S31: Collect and establish a laboratory data set of bonding pressure and acoustic emission multi-dimensional acoustic wave information, establish a bonding quality grade, assign grades 1 and 2 to high and low bonding strengths respectively, use the bonding quality grade of the bonded copper sheet as a reference signal, and annotate the collected data with the bonding quality grade; S32: Construct a transfer learning neural network model, including a time domain feature processing branch and a frequency domain feature processing branch. The time domain features are processed by a recurrent neural network, and the frequency domain features are processed by a convolutional neural network. The features extracted by the two branches are fused and connected to a fully connected layer. The bond head pressure signal collected by the bond head pressure sensor is used as input, multi-dimensional acoustic wave information is used as output, and the bond quality level is used as a label for training. S33: Verify and test the optimized transfer learning neural network model on laboratory datasets and online datasets, and fine-tune the parameters of the transfer learning neural network model on the online dataset to achieve knowledge transfer from the laboratory environment to the production environment.

5. The acoustic emission online evaluation method for copper sheet bonding strength according to claim 1, characterized in that: The bond mass fluctuation rate calculation in step S4 adopts a recursive prediction method, which specifically includes: The bond quality prediction value is calculated based on the current moment and historical data. The current prediction value is determined by the historical bond quality mean, the bond quality value of the corresponding sample at the previous moment, and the quality change of the newly added sample. The influence of historical data and new data is balanced by an adaptive weight coefficient. The fluctuation rate was calculated based on the predicted bonding mass values ​​of multiple consecutive samples. When the fluctuation rate was lower than 10%, it was determined to be a stable region; when the fluctuation rate was between 10% and 20%, it was determined to be a fluctuating region; and when the fluctuation rate was higher than 20%, it was determined to be an unstable region.

6. The acoustic emission online evaluation method for copper sheet bonding strength according to claim 1, characterized in that: The bonding pressure adjustment method in step S5 is: Standardize the collected bond head pressure and vertical acoustic wave information of the bonding position; Calculate the Pearson correlation coefficient between the normalized vertical acoustic wave information characteristics and the bond head pressure; Sort the acoustic wave information features vertical to the bonding position in descending order of correlation coefficient, and take the first n acoustic wave information feature values ​​to construct a subset; The relationship between the bond head pressure and each eigenvalue corresponding to the subset is calculated using the least squares method; Taking the minimum standard deviation of the overall bond strength in the bonding area as the objective function, the optimal bond head pressure is output; When the bonding quality is in the fluctuating zone, adjust the bonding pressure by 5% to 7%; when the bonding quality is in the unstable zone, adjust the bonding pressure by 8% to 10% until the bonding quality reaches the stable zone.

7. The acoustic emission online evaluation method for copper sheet bonding strength according to claim 1, characterized in that: The first acoustic emission sensor, the second acoustic emission sensor and the acoustic emission sensing array use different types of sensors according to the collection position and signal characteristics, including piezoelectric ceramic sensors for high-frequency signal collection, piezoelectric film sensors for wide-band signal collection and piezoelectric cantilever beam sensors for low-frequency signal collection.

8. The acoustic emission online evaluation method for copper sheet bonding strength according to claim 1, characterized in that: The first acoustic emission sensors are arranged in 3 to 12 numbers, the second acoustic emission sensors are arranged in 3 to 12 numbers, and the acoustic emission sensor array is arranged in 30 to 120 numbers. The spatial resolution and data redundancy of acoustic emission signal acquisition are improved through the high-density sensor array.

9. The acoustic emission online evaluation method for copper sheet bonding strength according to claim 1, characterized in that: After step S5 is completed, the method further includes: outputting the adjusted bond head pressure corresponding to the pressure to the bonding device; After performing the bonding process, verify the bond quality; Feedback the verification results to the bond strength evaluation model to achieve continuous optimization of the model; Establish a historical database of bonding parameters and bonding quality for continuous improvement of subsequent bonding processes.

10. An acoustic emission online evaluation system for copper sheet bonding strength, used to implement the acoustic emission online evaluation method for copper sheet bonding strength according to any one of claims 1 to 9, characterized in that: include: The data acquisition module includes a first acoustic emission sensor arranged above the bonding copper sheet mold, a second acoustic emission sensor arranged below the bonding copper sheet mold, an acoustic emission sensor array arranged below the bonding head at a position corresponding to the bonding chip position, and a bonding head pressure sensor. The first acoustic emission sensor and the second acoustic emission sensor are respectively used to collect longitudinal and transverse acoustic wave information of the bonding position. The acoustic emission sensor array is used to collect vertical acoustic wave information of the bonding position. The acoustic emission sensor array is spaced apart from the bonding copper sheet and does not contact the bonding copper sheet or the bonding copper sheet mold during the bonding process. Feature processing module: used to divide the bonding process into heating stage, pressing stage and pressure holding stage, respectively extract the time domain features, frequency domain features and time-frequency cross features of each stage, standardize the extracted features, and select the optimal feature subset by combining recursive feature elimination method with correlation calculation; Model building module: This module uses a transfer learning strategy to construct a laboratory dataset using data collected by the bond head pressure sensor and multi-dimensional acoustic wave information for model training. It then performs fine-tuning on the online dataset to form a bond strength assessment model. Quality assessment module: used to calculate the bond quality fluctuation rate based on the bond strength assessment model, determine the bond quality stability based on the fluctuation rate, and judge whether the bond quality is in the stable zone, fluctuation zone or unstable zone; Pressure adjustment module: used to optimize the bonding pressure parameters through the standard deviation minimization objective function. According to the correlation between the vertical acoustic wave information characteristics of the bonding position and the bond head pressure, the standard deviation of the bond strength at each position in the bonding area is calculated. The optimal bond head pressure is output with the goal of minimizing the overall standard deviation of the bond strength in the bonding area. The corresponding pressure adjustment strategy is executed according to the area where the bonding quality is located.

Images