Online temperature detection for ultrasonic welding joint strength and surface damage prediction method

Abstract

The present application provides an ultrasonic welding joint strength and surface damage prediction method for online temperature detection, adopts a K-type thermocouple to collect the peak temperature in the ultrasonic welding process online, and combines the process parameters of welding pressure, welding amplitude, welding time and welding power to construct a feature data set; after the collected data is preprocessed, a machine learning prediction model is used, and the lap shear load and the surface embedded damage depth are taken as the prediction targets to simultaneously predict the joint mechanical properties and the surface damage state; the present application can solve the problems of detection lag, high experimental cost and difficulty in simultaneously considering joint strength and surface damage evaluation in the existing ultrasonic welding quality evaluation process.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic welding quality monitoring and prediction technology, and in particular to a method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection. Specifically, it is a method for simultaneously predicting the strength and surface damage of ultrasonic welded joints by integrating online temperature detection. By acquiring temperature information of the welding process online and combining it with process parameters, a joint performance prediction model is established to achieve online evaluation and prediction of ultrasonic welding quality. Background Technology

[0002] Ultrasonic welding is a joining process that uses high-frequency vibration energy to induce plastic deformation and localized temperature rise at the welding interface, achieving solid-state fusion of materials. With its advantages of low temperature, high efficiency, no consumables, and environmental friendliness, it has been widely used in the chip packaging field. In the chip packaging process, chip pins and pins need to be precisely connected to components such as copper-clad ceramic substrates and lead frames. As a core interconnection process, ultrasonic welding directly determines the chip's conductivity, heat dissipation efficiency, and service reliability. It is a key step in the packaging of power semiconductor modules such as IGBTs and SiC, as well as consumer electronics chips. Currently, ultrasonic welding accounts for over 90% of all chip interconnection methods, with approximately 90% using gold wire ultrasonic welding and the remainder using aluminum wire and other precious metal wedge welding processes. It is widely used in fields that rely on high-precision chip packaging, such as automotive electronics, photovoltaic energy storage, and 5G communications.

[0003] While ultrasonic welding offers significant advantages in chip packaging, the precision requirements of chip welding are extremely high. The tiny size of chip leads means that even minor defects and quality fluctuations in the weld joint can lead to poor chip contact, heat dissipation failure, or even complete chip scrapping. Therefore, establishing a predictive method that integrates welding process status information and process parameter information to achieve simultaneous evaluation of joint strength and surface damage is crucial for improving ultrasonic welding quality control and reducing production costs. However, existing research primarily focuses on single-objective predictions based on process parameters, lacking effective utilization of welding process temperature information and failing to adequately assess the synergistic effects of joint strength and surface damage. Summary of the Invention

[0004] This invention proposes a method for predicting the strength and surface damage of ultrasonic welded joints based on online temperature detection, which can solve the problems of detection lag, high experimental costs, and difficulty in simultaneously evaluating joint strength and surface damage in existing ultrasonic welding quality evaluation processes.

[0005] The present invention adopts the following technical solution.

[0006] A method for predicting the strength and surface damage of ultrasonic welded joints based on online temperature detection is proposed. This method integrates online temperature detection for simultaneous prediction of the strength and surface damage of ultrasonic welded joints. It employs a K-type thermocouple to collect peak temperatures during the ultrasonic welding process online, and constructs a feature dataset by combining process parameters such as welding pressure, welding amplitude, welding time, and welding power. After preprocessing the collected data, a machine learning prediction model is used, with lap shear load and surface embedded damage depth as prediction targets, to achieve simultaneous prediction of the joint's mechanical properties and surface damage state.

[0007] The method includes the following steps;

[0008] Step 1: Use a temperature detection device to collect temperature data during the ultrasonic welding process online, and extract the peak temperature of the welding process; and simultaneously acquire the welding pressure, welding amplitude, welding time and welding power process parameters under the corresponding welding conditions;

[0009] Step 2: Perform outlier removal and normalization on the peak temperature and process parameters to construct a welding feature dataset;

[0010] Step 3: Using welding pressure, welding amplitude, welding time, welding power and peak temperature as input features, and joint strength characterized by lap shear load and surface embedded damage depth as prediction targets, establish a machine learning prediction model, namely a joint strength and surface damage synchronous prediction model.

[0011] Step 4: Use the acquired welding sample data to train the synchronous prediction model and establish the mapping relationship between the input features and the prediction target; after the model training is completed, input the feature data corresponding to the welding condition to be predicted into the model and output the joint strength and surface embedded damage prediction results under the corresponding condition.

[0012] In step 1, the temperature detection device is a type K thermocouple, which is placed near the welding area to collect temperature data during the welding process in real time and extract the peak temperature.

[0013] The K-type thermocouple has a temperature measurement range of -73°C to 482°C. During ultrasonic welding, the thermocouple probe is fixed between the welding copper sheets to directly collect the temperature change of the welding area and extract the highest temperature in the welding cycle corresponding to each welding condition as the peak temperature T_peak.

[0014] During ultrasonic welding, the process parameters can be accurately controlled by the ultrasonic welding equipment, including four core process parameters: welding pressure P, welding time t, welding amplitude A, and welding power W.

[0015] See Figure 1As shown, the following steps are used when acquiring the peak temperature of the ultrasonic welding process online in step 1:

[0016] Step 1: Wind the K-type thermocouple resistance wire to form a current path, connect it to the data acquisition unit, and ensure that the temperature measurement path is conductive and the data transmission is normal.

[0017] Step II: Place the T2 copper base plate in the ultrasonic welding machine, place the connected thermocouple resistance wire in the center of the welding area of ​​the base plate, and then cover the upper substrate on top of the resistance wire to ensure that the resistance wire is completely within the welding area and is accurately positioned.

[0018] Step III: Adjust the core process parameters of the ultrasonic welding machine, and adjust the welding pressure, welding time, and welding amplitude to the preset target values. After parameter calibration, start the ultrasonic welding machine to perform welding operations.

[0019] In this embodiment, the welding parameter range is as follows:

[0020] Table 2 Welding Parameter Range

[0021]

[0022] Step IV: During the welding process, the temperature change curve of the welding area is collected and recorded in real time by data acquisition equipment. The highest value within the welding cycle is extracted from the curve, which is the welding peak temperature. The peak temperature is the highest temperature value collected during the welding process, and its measurement range is 100℃~500℃.

[0023] Step V: After welding is completed, the mechanical properties of the welded joint are tested using a universal mechanical testing machine to determine its lap shear load; the surface of the welded joint is scanned and inspected using a white light interferometer to obtain the depth of surface damage caused by the knurling of the weld head.

[0024] In step 2, the data processing module organizes and filters all the online peak temperature data and corresponding process parameters, eliminates abnormal data interference and unifies the data units, and outputs standardized feature vectors. Data preprocessing provides high-quality input for subsequent batch dataset construction and machine learning model training.

[0025] The data preprocessing includes outlier removal and normalization. The normalization process uses the Min-Max normalization method to map each input feature to a unified numerical range; that is, it maps the feature values ​​to the 0-1 range to eliminate the influence of different dimensions between parameters and ensure the comparability of the data.

[0026] In step 3, the batch dataset prepared in step 2 is loaded through the neural network training module, and the training of the lap shear load and surface damage prediction model is completed. After the training is qualified, it is used to predict the results of the untested working conditions.

[0027] The synchronous prediction model is a machine learning model, including one or more of the following: neural network model, support vector machine model, random forest model, or gradient boosting tree model;

[0028] Machine learning models can predict joint strength quality or surface embedded damage by changing different outputs without altering the structure.

[0029] In step 3, the electron backscatter diffraction (EBSD) image of the welded joint is acquired, and the feature vector characterizing the microstructure of the welded joint is extracted by a deep learning model. The feature vector, along with the welding process parameters and peak temperature, is then input into the machine learning model for prediction.

[0030] The deep learning model is a variational autoencoder (VAE), which compresses the EBSD image into a low-dimensional latent space through the encoder to obtain latent variable features for characterizing the microstructure of the weld joint.

[0031] The EBSD image includes one or more of the following: inverse pole figure (IPF), kernel average mismatch figure (KAM), or grain boundary figure.

[0032] The method for obtaining EBSD images of welded joints as microstructural characterization information includes the following steps;

[0033] Step AⅠ: Acquire EBSD images of the welded joint and adjust the image size and format accordingly;

[0034] Step AⅡ: Construct a variational autoencoder (VAE) network to compress and encode the EBSD image to obtain a low-dimensional latent variable feature vector; the latent variable feature vector has a dimension of 4 to 64, preferably 8 or 16.

[0035] Step AⅢ: The obtained latent variable features are fused with welding pressure, welding time, welding amplitude, welding power and peak temperature to construct an extended input feature set;

[0036] Step AⅣ: Input the extended feature set into the machine learning model for training to achieve prediction of lap shear load and surface embedded damage.

[0037] Experimental results show that by introducing EBSD image features, the model can further characterize the microstructure of the welded joint and improve the predictive ability of the model to predict the joint performance.

[0038] The synchronous prediction model is constructed as follows: First, a three-layer hidden-layer feedforward neural network is built based on Matlab 2023b. The network structure is set as follows: input layer (5 neurons) → first hidden layer (32 neurons) → second hidden layer (16 neurons) → third hidden layer (8 neurons) → output layer (1 neuron). All three hidden layers use the sigmoid activation function to achieve a non-linear mapping between input features and the prediction target, accurately capturing the complex coupling relationship between welding parameters, lap shear load, and surface damage depth. The output layer uses the purelin linear function to adapt to the prediction requirements of continuous lap shear load values ​​and surface damage depth values. The model training parameters are set, the training algorithm is the trainlm algorithm, the learning rate is set to 0.01, the maximum number of training iterations is set to 1000, and the training target error is set to... This ensures the accuracy and efficiency of model training;

[0039] The standardized welding dataset is then divided into training and testing sets using a 5-fold cross-validation scheme. The divided datasets are then input into the constructed neural network. The lap shear load and surface embedding depth are set as the model prediction target values, respectively. No other parameters of the neural network need to be changed. The model learns and trains for the two prediction targets in sequence, achieving accurate prediction of two indicators by the same network.

[0040] Then, the predictive performance of the trained neural network model is quantitatively evaluated to verify the model's prediction accuracy for lap shear load and surface damage depth, and to determine whether the model can be used for prediction of untested working conditions.

[0041] The specific steps for quantitatively evaluating the predictive performance of a neural network model are as follows:

[0042] Step BⅠ: Input the test set data obtained from cross-validation into the neural network model that has completed dual-objective training to obtain the model prediction values ​​of lap shear load and surface damage depth, respectively.

[0043] Step BⅡ: Select the root mean square error (RMSE) and the coefficient of determination (…). As the core quantitative evaluation index, the values ​​of two indicators—overlap shear load and surface damage depth—are calculated separately for the model's prediction results. The calculation formulas are as follows:

[0044] Root mean square error:

[0045] Coefficient of determination:

[0046] in, The number of samples in the test set. These are measured values. For predicted values, This represents the average of the measured values.

[0047] Step BⅢ: Set the model performance pass / fail threshold: RMSE is less than the preset error threshold, and If the quantitative indicators of the two predicted results, namely the lap shear load and the surface damage depth, both meet the qualified threshold requirements, the prediction performance of the neural network model is deemed qualified and can be directly used to predict the lap shear load and surface damage depth in ultrasonic welding conditions that have not been tested. If the threshold requirements are not met, the network parameters should be readjusted or the dataset should be expanded for model retraining.

[0048] like Figure 3 and Figure 4 As shown, the trained neural network achieves a prediction coefficient of 0.9556 for the lap shear load of the joint. Figure 3 (a) The coefficient of determination for predicting surface embedded damage reaches 0.9401 ( Figure 3 (b)). Figure 4 The impact of peak temperature characteristics on model prediction performance is demonstrated. Results show that in joint strength prediction, the introduction of peak temperature has limited effect on improving model performance; however, in surface embedded damage prediction, the introduction of peak temperature significantly improves the model's coefficient of determination and further reduces the prediction error. This indicates that peak temperature can effectively reflect the heat input state during welding and has a strong characterization ability for the surface damage formation process, thereby improving the model's prediction accuracy for surface damage.

[0049] This invention acquires the peak welding temperature online and constructs a feature space by combining it with process parameters such as welding pressure, welding amplitude, welding time, and welding power. It then uses a machine learning model to establish a mapping relationship between process parameters and joint performance, enabling simultaneous prediction of lap shear load and surface embedded damage. This provides a basis for welding process optimization and online quality control.

[0050] Compared with existing technologies, this invention integrates welding process temperature information with process parameters, which improves the model's ability to characterize the welding state. It can quickly predict the performance of joints under untested conditions while reducing the number of experiments, providing technical support for ultrasonic welding process optimization and online quality control.

[0051] The technical solution described in this invention differs from existing technical approaches that improve prediction performance by optimizing detection equipment or training a single model in the following ways: This invention uses K-type thermocouples to collect peak welding temperatures in batches online, combines this data with corresponding core process parameters to form a batch dataset, and uses the same batch of data to simultaneously train a dual-prediction target model. This accurately uncovers the correlation between welding parameters and prediction targets, eliminating the need to conduct experiments for every welding condition. The trained model can predict results that have not been tested, significantly reducing experimental workload and costs while improving prediction accuracy. This provides data support for optimizing precision ultrasonic welding conditions for chips. Attached Figure Description

[0052] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0053] Appendix Figure 1 This is a schematic diagram of the online temperature detection scheme used in Example 1 of the present invention;

[0054] Appendix Figure 2 This is a schematic diagram of the neural network structure optimized using a genetic algorithm established in Example 2 of the present invention;

[0055] Appendix Figure 3 This is a schematic diagram illustrating the predictive effect of the present invention on lap shear load and surface embedded damage;

[0056] Appendix Figure 4 This is a schematic diagram illustrating the effect of the welding peak temperature input in this invention on the prediction effect;

[0057] Appendix Figure 5 This is a schematic diagram of obtaining EBSD images of welded joints as microstructure characterization information in an embodiment of the present invention;

[0058] Figure 1 In the middle: 1-Ultrasonic welding machine; 2-Upper substrate; 3-K-type thermocouple resistance wire; 4-Copper base plate; 5-Welding area. Detailed Implementation

[0059] The following examples are used to illustrate the present invention, but are not intended to limit the scope of the invention.

[0060] As shown in the figure, an online temperature detection method for predicting the strength and surface damage of ultrasonic welded joints is presented. This method integrates online temperature detection for simultaneous prediction of the strength and surface damage of ultrasonic welded joints. A K-type thermocouple is used to collect the peak temperature during the ultrasonic welding process online, and a feature dataset is constructed by combining process parameters such as welding pressure, welding amplitude, welding time, and welding power. After preprocessing the collected data, a machine learning prediction model is used, with lap shear load and surface embedded damage depth as prediction targets, to achieve simultaneous prediction of the joint's mechanical properties and surface damage state.

[0061] The method includes the following steps;

[0062] Step 1: Use a temperature detection device to collect temperature data during the ultrasonic welding process online, and extract the peak temperature of the welding process; and simultaneously acquire the welding pressure, welding amplitude, welding time and welding power process parameters under the corresponding welding conditions;

[0063] Step 2: Perform outlier removal and normalization on the peak temperature and process parameters to construct a welding feature dataset;

[0064] Step 3: Using welding pressure, welding amplitude, welding time, welding power and peak temperature as input features, and joint strength characterized by lap shear load and surface embedded damage depth as prediction targets, establish a machine learning prediction model, namely a joint strength and surface damage synchronous prediction model.

[0065] Step 4: Use the acquired welding sample data to train the synchronous prediction model and establish the mapping relationship between the input features and the prediction target; after the model training is completed, input the feature data corresponding to the welding condition to be predicted into the model and output the joint strength and surface embedded damage prediction results under the corresponding condition.

[0066] In step 1, the temperature detection device is a type K thermocouple, which is placed near the welding area to collect temperature data during the welding process in real time and extract the peak temperature.

[0067] The K-type thermocouple has a temperature measurement range of -73°C to 482°C. During ultrasonic welding, the thermocouple probe is fixed between the welding copper sheets to directly collect the temperature change of the welding area and extract the highest temperature in the welding cycle corresponding to each welding condition as the peak temperature T_peak.

[0068] During ultrasonic welding, the process parameters can be accurately controlled by the ultrasonic welding equipment, including four core process parameters: welding pressure P, welding time t, welding amplitude A, and welding power W.

[0069] See Figure 1 As shown, the following steps are used when acquiring the peak temperature of the ultrasonic welding process online in step 1:

[0070] Step 1: Wind the K-type thermocouple resistance wire to form a current path, connect it to the data acquisition unit, and ensure that the temperature measurement path is conductive and the data transmission is normal.

[0071] Step II: Place the T2 copper base plate in the ultrasonic welding machine, place the connected thermocouple resistance wire in the center of the welding area of ​​the base plate, and then cover the upper substrate on top of the resistance wire to ensure that the resistance wire is completely within the welding area and is accurately positioned.

[0072] Step III: Adjust the core process parameters of the ultrasonic welding machine, and adjust the welding pressure, welding time, and welding amplitude to the preset target values. After parameter calibration, start the ultrasonic welding machine to perform welding operations.

[0073] In this embodiment, the welding parameter range is as follows:

[0074] Table 2 Welding Parameter Range

[0075]

[0076] Step IV: During the welding process, the temperature change curve of the welding area is collected and recorded in real time by data acquisition equipment. The highest value within the welding cycle is extracted from the curve, which is the welding peak temperature. The peak temperature is the highest temperature value collected during the welding process, and its measurement range is 100℃~500℃.

[0077] Step V: After welding is completed, the mechanical properties of the welded joint are tested using a universal mechanical testing machine to determine its lap shear load; the surface of the welded joint is scanned and inspected using a white light interferometer to obtain the depth of surface damage caused by the knurling of the weld head.

[0078] In step 2, the data processing module organizes and filters all the online peak temperature data and corresponding process parameters, eliminates abnormal data interference and unifies the data units, and outputs standardized feature vectors. Data preprocessing provides high-quality input for subsequent batch dataset construction and machine learning model training.

[0079] The data preprocessing includes outlier removal and normalization. The normalization process uses the Min-Max normalization method to map each input feature to a unified numerical range; that is, it maps the feature values ​​to the 0-1 range to eliminate the influence of different dimensions between parameters and ensure the comparability of the data.

[0080] In step 3, the batch dataset prepared in step 2 is loaded through the neural network training module, and the training of the lap shear load and surface damage prediction model is completed. After the training is qualified, it is used to predict the results of the untested working conditions.

[0081] The synchronous prediction model is a machine learning model, including one or more of the following: neural network model, support vector machine model, random forest model, or gradient boosting tree model;

[0082] Machine learning models can predict joint strength quality or surface embedded damage by changing different outputs without altering the structure.

[0083] In step 3, the electron backscatter diffraction (EBSD) image of the welded joint is acquired, and the feature vector characterizing the microstructure of the welded joint is extracted by a deep learning model. The feature vector, along with the welding process parameters and peak temperature, is then input into the machine learning model for prediction.

[0084] The deep learning model is a variational autoencoder (VAE), which compresses the EBSD image into a low-dimensional latent space through the encoder to obtain latent variable features for characterizing the microstructure of the weld joint.

[0085] The EBSD image includes one or more of the following: inverse pole figure (IPF), kernel average mismatch figure (KAM), or grain boundary figure.

[0086] The method for obtaining EBSD images of welded joints as microstructural characterization information includes the following steps;

[0087] Step AⅠ: Acquire EBSD images of the welded joint and adjust the image size and format accordingly;

[0088] Step AⅡ: Construct a variational autoencoder (VAE) network to compress and encode the EBSD image to obtain a low-dimensional latent variable feature vector; the latent variable feature vector has a dimension of 4 to 64, preferably 8 or 16.

[0089] Step AⅢ: The obtained latent variable features are fused with welding pressure, welding time, welding amplitude, welding power and peak temperature to construct an extended input feature set;

[0090] Step AⅣ: Input the extended feature set into the machine learning model for training to achieve prediction of lap shear load and surface embedded damage.

[0091] Experimental results show that by introducing EBSD image features, the model can further characterize the microstructure of the welded joint and improve the predictive ability of the model to predict the joint performance.

[0092] The synchronous prediction model is constructed as follows: First, a three-layer hidden-layer feedforward neural network is built based on Matlab 2023b. The network structure is set as follows: input layer (5 neurons) → first hidden layer (32 neurons) → second hidden layer (16 neurons) → third hidden layer (8 neurons) → output layer (1 neuron). All three hidden layers use the sigmoid activation function to achieve a non-linear mapping between input features and the prediction target, accurately capturing the complex coupling relationship between welding parameters, lap shear load, and surface damage depth. The output layer uses the purelin linear function to adapt to the prediction requirements of continuous lap shear load values ​​and surface damage depth values. The model training parameters are set, the training algorithm is the trainlm algorithm, the learning rate is set to 0.01, the maximum number of training iterations is set to 1000, and the training target error is set to... This ensures the accuracy and efficiency of model training;

[0093] The standardized welding dataset is then divided into training and testing sets using a 5-fold cross-validation scheme. The divided datasets are then input into the constructed neural network. The lap shear load and surface embedding depth are set as the model prediction target values, respectively. No other parameters of the neural network need to be changed. The model learns and trains for the two prediction targets in sequence, achieving accurate prediction of two indicators by the same network.

[0094] Then, the predictive performance of the trained neural network model is quantitatively evaluated to verify the model's prediction accuracy for lap shear load and surface damage depth, and to determine whether the model can be used for prediction of untested working conditions.

[0095] The specific steps for quantitatively evaluating the predictive performance of a neural network model are as follows:

[0096] Step BⅠ: Input the test set data obtained from cross-validation into the neural network model that has completed dual-objective training to obtain the model prediction values ​​of lap shear load and surface damage depth, respectively.

[0097] Step BⅡ: Select the root mean square error (RMSE) and the coefficient of determination (…). As the core quantitative evaluation index, the values ​​of two indicators—overlap shear load and surface damage depth—are calculated separately for the model's prediction results. The calculation formulas are as follows:

[0098] Root mean square error:

[0099] Coefficient of determination:

[0100] in, The number of samples in the test set. These are measured values. For predicted values, This represents the average of the measured values.

[0101] Step BⅢ: Set the model performance pass / fail threshold: RMSE is less than the preset error threshold, and If the quantitative indicators of the two predicted results, namely the lap shear load and the surface damage depth, both meet the qualified threshold requirements, the prediction performance of the neural network model is deemed qualified and can be directly used to predict the lap shear load and surface damage depth in ultrasonic welding conditions that have not been tested. If the threshold requirements are not met, the network parameters should be readjusted or the dataset should be expanded for model retraining.

[0102] like Figure 3 and Figure 4 As shown, the trained neural network achieves a prediction coefficient of 0.9556 for the lap shear load of the joint. Figure 3 (a) The coefficient of determination for predicting surface embedded damage reaches 0.9401 ( Figure 3 (b)). Figure 4 The impact of peak temperature characteristics on model prediction performance is demonstrated. Results show that in joint strength prediction, the introduction of peak temperature has limited effect on improving model performance; however, in surface embedded damage prediction, the introduction of peak temperature significantly improves the model's coefficient of determination and further reduces the prediction error. This indicates that peak temperature can effectively reflect the heat input state during welding and has a strong characterization ability for the surface damage formation process, thereby improving the model's prediction accuracy for surface damage.

[0103] Example 1: Online temperature acquisition and connector performance testing

[0104]

[0105] See Figure 1 As shown, the following steps are required to acquire the peak temperature during the ultrasonic welding process online:

[0106] Ⅰ. Wind the K-type thermocouple resistance wire to form a current path, connect it to the data acquisition unit, and ensure that the temperature measurement path is conductive and the data transmission is normal;

[0107] II. Place a T2 copper base plate in the ultrasonic welding machine, place the connected thermocouple resistance wire in the center of the welding area of ​​the base plate, and then cover the upper substrate on top of the resistance wire to ensure that the resistance wire is completely within the welding area and is accurately positioned.

[0108] III. Adjust the core process parameters of the ultrasonic welding machine, and adjust the welding pressure, welding time, and welding amplitude to the preset target values. After parameter calibration, start the ultrasonic welding machine to perform welding operations.

[0109] IV. During the welding process, the temperature change curve of the welding area is collected and recorded in real time by data acquisition equipment. The highest value within the welding cycle is extracted from the curve, which is the welding peak temperature.

[0110] V. After welding, the mechanical properties of the welded joint are tested using a universal mechanical testing machine to determine its lap shear load; the surface of the welded joint is scanned and inspected using a white light interferometer to obtain the depth of surface damage caused by the knurling of the weld head.

[0111] Example 2: Neural Network Training

[0112] See Figure 2 As shown, the neural network is trained using collected welding-related data to achieve accurate prediction of lap shear load and surface damage depth. The specific steps are as follows:

[0113] I. The collected welding process parameters, peak temperatures and corresponding performance test data are preprocessed. The Min-Max standardization scheme is used to standardize all feature values, and the feature values ​​are uniformly mapped to the 0-1 interval to eliminate the influence of different dimensions between parameters and ensure the comparability of data.

[0114] II. A three-layer hidden-layer feedforward neural network was built based on Matlab 2023b. The network structure was set as follows: input layer (5 neurons) → first hidden layer (32 neurons) → second hidden layer (16 neurons) → third hidden layer (8 neurons) → output layer (1 neuron). All three hidden layers used the sigmoid activation function to achieve a non-linear mapping between input features and the prediction target, accurately capturing the complex coupling relationship between welding parameters, lap shear load, and surface damage depth. The output layer used the purelin linear function to adapt to the prediction requirements of continuous lap shear load values ​​and surface damage depth values. The model training parameters were set, using the trainlm algorithm with a learning rate of 0.01, a maximum training iteration count of 1000, and a training target error of 1e-5 to ensure the accuracy and efficiency of model training.

[0115] III. The standardized welding dataset is divided into training and testing sets using a 5-fold cross-validation scheme. The divided datasets are then input into the constructed neural network. The lap shear load and surface embedding depth are set as the model's prediction target values, respectively. No other parameters of the neural network need to be changed. The model learns and trains for the two prediction targets in sequence, achieving accurate prediction of two indicators by the same network.

[0116] Example 3: Evaluation of Neural Network Prediction Performance

[0117] The predictive performance of the neural network model trained in Example 2 was quantitatively evaluated to verify the model's prediction accuracy for lap shear load and surface damage depth, and to determine whether the model can be used for prediction of untested working conditions.

[0118] Example 4: Construction of a Prediction Model Based on EBSD Image Features

[0119] See Figure 5 As shown, in order to further improve the model's ability to predict the performance of welded joints, EBSD images of the welded joints were obtained as microstructure characterization information.

[0120] Ⅰ. Acquire EBSD images of the welded joints and uniformly adjust the image size and format;

[0121] II. Construct a variational autoencoder (VAE) network to compress and encode the EBSD image to obtain a low-dimensional latent variable feature vector; the latent variable feature vector has a dimension of 4 to 64, preferably 8 or 16.

[0122] III. The obtained latent variable features are fused with welding pressure, welding time, welding amplitude, welding power, and peak temperature to construct an extended input feature set;

[0123] IV. Input the extended feature set into the machine learning model for training to achieve prediction of lap shear load and surface embedded damage.

[0124] Experimental results show that by introducing EBSD image features, the model can further characterize the microstructure of the welded joint and improve the predictive ability of the model to predict the joint performance.

Claims

1. A method for predicting the strength and surface damage of ultrasonic welded joints based on online temperature detection, characterized in that: The method uses a K-type thermocouple to collect the peak temperature during the ultrasonic welding process online, and constructs a feature dataset by combining process parameters such as welding pressure, welding amplitude, welding time, and welding power. After preprocessing the collected data, a machine learning prediction model is used, with lap shear load and surface embedded damage depth as prediction targets, to simultaneously predict the mechanical properties and surface damage state of the joint.

2. The method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection according to claim 1, characterized in that: The method includes the following steps; Step 1: Use a temperature detection device to collect temperature data during the ultrasonic welding process online, and extract the peak temperature of the welding process; and simultaneously acquire the welding pressure, welding amplitude, welding time and welding power process parameters under the corresponding welding conditions; Step 2: Perform outlier removal and normalization on the peak temperature and process parameters to construct a welding feature dataset; Step 3: Using welding pressure, welding amplitude, welding time, welding power and peak temperature as input features, and joint strength characterized by lap shear load and surface embedded damage depth as prediction targets, establish a machine learning prediction model, namely a joint strength and surface damage synchronous prediction model. Step 4: Use the acquired welding sample data to train the synchronous prediction model and establish the mapping relationship between the input features and the prediction target; after the model training is completed, input the feature data corresponding to the welding condition to be predicted into the model and output the joint strength and surface embedded damage prediction results under the corresponding condition.

3. The method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection according to claim 2, characterized in that: In step 1, the temperature detection device is a type K thermocouple, which is placed near the welding area to collect temperature data during the welding process in real time and extract the peak temperature. The K-type thermocouple has a temperature measurement range of -73°C to 482°C. During ultrasonic welding, the thermocouple probe is fixed between the welding copper sheets to directly collect the temperature change of the welding area and extract the highest temperature in the welding cycle corresponding to each welding condition as the peak temperature T_peak. During ultrasonic welding, process parameters are controlled by ultrasonic welding equipment, including welding pressure P, welding time t, welding amplitude A, and welding power W.

4. The method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection according to claim 3, characterized in that: The following steps are used when acquiring the peak temperature of the ultrasonic welding process online in step 1: Step 1: Wind the K-type thermocouple resistance wire to form a current path, connect it to the data acquisition unit, and ensure that the temperature measurement path is conductive and the data transmission is normal. Step II: Place the T2 copper base plate in the ultrasonic welding machine, place the connected thermocouple resistance wire in the center of the welding area of ​​the base plate, and then cover the upper substrate on top of the resistance wire to ensure that the resistance wire is completely within the welding area and is accurately positioned. Step 3: Adjust the core process parameters of the ultrasonic welding machine, and adjust the welding pressure, welding time, and welding amplitude to the preset target values. After parameter calibration, start the ultrasonic welding machine to perform welding operations. Step IV: During the welding process, the temperature change curve of the welding area is collected and recorded in real time by data acquisition equipment. The highest value within the welding cycle is extracted from the curve, which is the welding peak temperature. The peak temperature is the highest temperature value collected during the welding process, and its measurement range is 100℃~500℃. Step V: After welding is completed, the mechanical properties of the welded joint are tested using a universal mechanical testing machine to determine its lap shear load; the surface of the welded joint is scanned and inspected using a white light interferometer to obtain the depth of surface damage caused by the knurling of the weld head.

5. The method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection according to claim 2, characterized in that: In step 2, the data processing module organizes and filters all the online peak temperature data and corresponding process parameters, eliminates abnormal data interference and unifies the data units, and outputs standardized feature vectors. Data preprocessing provides high-quality input for subsequent batch dataset construction and machine learning model training. The data preprocessing includes outlier removal and normalization, wherein the normalization process uses the Min-Max normalization method to map each input feature to a uniform numerical range.

6. The method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection according to claim 2, characterized in that: In step 3, the batch dataset prepared in step 2 is loaded through the neural network training module, and the training of the lap shear load and surface damage prediction model is completed. After the training is qualified, it is used to predict the results of the untested working conditions. The synchronous prediction model is a machine learning model, including a neural network model, a support vector machine model, a random forest model, or a gradient boosting tree model. Machine learning models can predict joint strength quality or surface embedded damage by varying different outputs.

7. The method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection according to claim 2, characterized in that: In step 3, the electron backscatter diffraction (EBSD) image of the welded joint is acquired, and the feature vector characterizing the microstructure of the welded joint is extracted by a deep learning model. The feature vector, along with the welding process parameters and peak temperature, is then input into the machine learning model for prediction. The deep learning model is a variational autoencoder, which compresses the EBSD image into a low-dimensional latent space through the encoder to obtain latent variable features for characterizing the microstructure of the welded joint. The EBSD images include inverse pole figures, kernel-average mismatch figures, or grain boundary figures.

8. The method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection according to claim 7, characterized in that: The method for obtaining EBSD images of welded joints as microstructural characterization information includes the following steps; Step AⅠ: Acquire EBSD images of the welded joint and adjust the image size and format accordingly; Step AⅡ: Construct a variational autoencoder (VAE) network to compress and encode the EBSD image, obtaining a low-dimensional latent variable feature vector; the latent variable feature vector has a dimension of 4 to 64. Step AⅢ: The obtained latent variable features are fused with welding pressure, welding time, welding amplitude, welding power and peak temperature to construct an extended input feature set; Step AⅣ: Input the extended feature set into the machine learning model for training to achieve prediction of lap shear load and surface embedded damage.

9. The method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection according to claim 7, characterized in that: The synchronous prediction model is constructed as follows: First, a three-layer hidden-layer feedforward neural network is built, with the network structure set as: input layer → first hidden layer → second hidden layer → third hidden layer → output layer; all three hidden layers use the sigmoid activation function to achieve a non-linear mapping between input features and prediction targets, accurately capturing the complex coupling relationship between welding parameters, lap shear load, and surface damage depth; the output layer uses the purelin linear function to adapt to the prediction requirements of continuous lap shear load values ​​and surface damage depth values; the model training parameters are set, the training algorithm uses the trainlm algorithm, and the training target error is set to... This ensures the accuracy and efficiency of model training; The standardized welding dataset is then divided into a training set and a test set. The divided dataset is then input into the constructed neural network. The lap shear load and surface embedding depth are set as the model prediction target values, respectively. The model learns and trains for the two prediction targets in turn, so that the same network can accurately predict the two indicators. Then, the predictive performance of the trained neural network model is quantitatively evaluated to verify the model's prediction accuracy for lap shear load and surface damage depth, and to determine whether the model can be used for prediction of untested working conditions.

10. The method for predicting the strength and surface damage of ultrasonic welded joints by online temperature detection according to claim 9, characterized in that: The specific steps for quantitatively evaluating the predictive performance of a neural network model are as follows: Step BⅠ: Input the test set data obtained from cross-validation into the neural network model that has completed dual-objective training to obtain the model prediction values ​​of lap shear load and surface damage depth, respectively. Step BⅡ: Select the root mean square error and coefficient of determination as the core quantitative evaluation indicators, and calculate the values ​​of the two indicators for the model's prediction results of lap shear load and surface damage depth, respectively. The calculation formulas are as follows: Root mean square error: Coefficient of determination: in, The number of samples in the test set. These are measured values. For predicted values, This represents the average of the measured values. Step BⅢ: Set the model performance pass / fail threshold: RMSE is less than the preset error threshold, and If the quantitative indicators of the two predicted results, namely the lap shear load and the surface damage depth, both meet the qualified threshold requirements, the prediction performance of the neural network model is deemed qualified and can be directly used to predict the lap shear load and surface damage depth in ultrasonic welding conditions that have not been tested. If the threshold requirements are not met, the network parameters should be readjusted or the dataset should be expanded for model retraining.

Images