Ultrasonic bonding machine and bonding method

Abstract

The invention discloses an ultrasonic bonding machine and a bonding method, and belongs to the field of ultrasonic welding. Wherein the ultrasonic bonding machine comprises a bonding head and a quality detection system; the bonding head comprises a detection device, and the detection device detects bonding parameters including ultrasonic current, ultrasonic voltage, ultrasonic power and deformation quantity; the quality detection system comprises a memory and a processor, a quality detection program is stored in the memory, the quality detection program can detect the bonding quality according to the bonding parameters when being executed by the processor, and the quality detection program is obtained based on given bonding data set training. The ultrasonic bonding machine can effectively improve the efficiency and accuracy of bonding quality detection and improve the economical efficiency of the bonding process.

Description

Technical Field

[0001] The invention belongs to the field of ultrasonic welding, and in particular relates to an ultrasonic bonding machine and a bonding method. Background Art

[0002] Bonding is an important process in the electronics manufacturing industry, and the quality of bonding has a significant impact on product quality. Effective monitoring of bonding quality during continuous production is of great significance to improving product economy and quality.

[0003] Patent CN116929277B discloses a monitoring mechanism for copper wire bonding, which uses a diameter gauge to detect the copper wire diameter, and detects the resistance change, deformation and color change of the copper wire in real time during the bonding process, and compares the change amplitude with the threshold determined by the deep learning algorithm to determine whether the bonding quality meets the requirements. The deep learning algorithm is based on a neural network model, predicts the bonding quality based on the copper wire resistance, deformation and color, and establishes a mapping relationship from copper wire parameters to bonding quality.

[0004] However, the monitoring mechanism uses a fixed threshold to control the bonding quality. The number of parameters is small, and the coupling relationship between the parameters is not sufficiently considered. When it is necessary to distinguish between cold solder joints, over solder joints, broken wires and normal states, it is difficult to make an accurate judgment only by the copper wire parameter threshold.

[0005] Therefore, providing an ultrasonic bonding machine that can accurately detect bonding quality has positive significance for improving bonding quality. Summary of the invention

[0006] The purpose of the present invention is to provide an ultrasonic bonding machine to improve the bonding quality. The present invention also provides a bonding method.

[0007] According to an embodiment of one aspect of the present invention, there is provided an ultrasonic bonding machine, the ultrasonic bonding machine comprising a bonding head and a quality detection system, wherein:

[0008] The bonding head includes a bracket, a working part, a pressure motor and a detection device.

[0009] The working part, the pressure motor and the detection device are connected to the bracket.

[0010] The pressure motor drives the working part to vibrate,

[0011] The detection device detects the bonding parameters of the bonding head during operation, and the bonding parameters include at least two of ultrasonic current, ultrasonic voltage, ultrasonic power, and deformation amount.

[0012] The quality detection system comprises a memory and a processor, wherein the memory stores a quality detection program and the bonding parameters detected by the detection device.

[0013] The processor can execute the quality detection program to detect the bonding quality of the bonding operation position corresponding to the bonding parameter.

[0014] The quality inspection procedure is obtained by:

[0015] Providing a bonding data set, the bonding data set comprising a bonding sample, the bonding sample comprising a bonding parameter group with a label, the label comprising normal, cold soldering, over soldering and broken wire;

[0016] A deep learning neural network model is established, the bond data set is divided into a training set and a test set, the training set is used for training, and the test set is used for testing until the accuracy of label classification of the test set is no less than a given threshold.

[0017] The ultrasonic bonding machine can effectively improve the judgment accuracy of bonding quality management, use multiple parameters to make multi-dimensional comprehensive judgments, effectively distinguish different production states such as normal bonding, cold welding, over-welding and broken wires, and improve continuous production efficiency and management level.

[0018] Furthermore, in some embodiments, the ultrasonic bonding machine also includes a pressure spring, and two of the pressure springs are arranged at both ends of the vibration direction of the pressure motor, so that the pressure motor is connected to the bracket through the pressure spring; the pressure spring is configured in a circular shape and is arranged coaxially with the pressure motor.

[0019] The purpose of designing the pressure spring into a circle is to ensure that the force on all sides of the bonding head is uniform during the deformation process, thereby ensuring that the splitting knife will not swing back and forth or left and right, and that the splitting knife is always perpendicular to the bonding surface.

[0020] Furthermore, in some embodiments, the pressure spring is provided with a plurality of internal fixing points and a plurality of external fixing points, the internal fixing points and the external fixing points are respectively arranged in a centrally symmetrical manner, the internal fixing points are connected to the pressure motor, and the external fixing points are connected to the bracket.

[0021] Furthermore, in some embodiments, the pressure spring sheet is provided with a plurality of through grooves, the through grooves penetrate the pressure spring sheet axially and are centrally symmetrically arranged, and each of the through grooves extends between one of the outer fixing points and one of the inner fixing points.

[0022] Further, in some embodiments, the through grooves are distributed in a spiral shape.

[0023] Furthermore, in some embodiments, the detection device includes a matching grating and a reading head, wherein:

[0024] The grating is fixedly arranged relative to the pressure motor, and the reading head is fixedly arranged relative to the bracket.

[0025] Furthermore, in some embodiments, the neural network model includes a ResNet50 deep learning neural network.

[0026] Furthermore, in some embodiments, the process of establishing a deep learning neural network is as follows:

[0027] Use Python to import the ResNet50 model, add a fully connected layer for classification, normalize the output of the fully connected layer, and obtain the probability distribution of four categories: cold soldering, over soldering, broken wire, and normal. Calculate the loss function based on the predicted results and the true labels, and update the model weight parameters through the back propagation algorithm;

[0028] Use transfer learning to load existing training weights;

[0029] The labels are defined, values ​​are assigned to cold soldering, over soldering, broken wire and normal respectively, and data is input;

[0030] The input data is passed to the input layer of the neural network, the activation value is calculated using the ReLU function, and it is passed to the next layer of neurons layer by layer through the inter-layer connections of the neural network until the output layer;

[0031] Calculating a loss function, wherein the loss function includes a cross entropy loss function;

[0032] Calculate the gradient of the loss function for each layer of neurons, and back-propagate to update the parameters of the neurons in the previous layer;

[0033] Iterative calculation is performed to minimize the loss function and complete the training.

[0034] According to another embodiment of the present invention, a bonding method is provided. The bonding method uses the ultrasonic bonding machine provided in any of the aforementioned embodiments.

[0035] Furthermore, in some embodiments, during the bonding process, the bonding data set is updated according to actual production data.

[0036] Furthermore, in some embodiments, the memory includes a cache device, and during the bonding process, the cache device stores the bonding parameters, and transmits the bonding parameters to the processor after a given number of bonding processes are performed. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of a local structure of a bonding machine in one embodiment;

[0038] Figure 2 for Figure 1 Schematic diagram of the AA section structure;

[0039] Figure 3 This is a schematic diagram of the structure of a pressure spring in an embodiment;

[0040] Figure 4 A schematic diagram of obtaining a grating signal in an embodiment;

[0041] Figure 5 This is a schematic diagram of a model training process in an embodiment;

[0042] Figure 6 A schematic diagram of forward propagation of a model in an embodiment;

[0043] Figure 7 A schematic diagram of a quality inspection procedure flow chart in one embodiment;

[0044] Figure 8a is a bonding power curve in a normal state in an embodiment;

[0045] Figure 8b is a deformation variable curve of a normal state in an embodiment;

[0046] Figure 9a is a bonding power curve in a disconnected state in one embodiment;

[0047] Figure 9b is a deformation amount curve of a disconnected state in one embodiment;

[0048] Fig.10a is a bonding power curve of an over-welding state in an embodiment;

[0049] Fig.10b is a deformation amount curve of an over-welding state in an embodiment;

[0050] Fig.11a is a bonding power curve of a cold solder joint in an embodiment;

[0051] Fig.11b In one embodiment, the deformation amount of the cold welding turret is wireless.

[0052] Meaning of the reference symbols:

[0053] 1-working part; 11-cutter; 12-cutter; 13-wire guide nozzle; 21-upper bracket; 22-lower bracket; 23-clamp motor; 24-transducer; 30-pressure motor; 31-pressure spring; 311-external fixing point; 312-internal fixing point; 313-through slot; 41-grating ruler; 42-reading head.

[0054] The purpose of the above drawings is to explain the present invention in detail so that those skilled in the art can understand the technical concept of the present invention, but it is not intended to limit the present invention. For the sake of simplicity, the above drawings only schematically illustrate the structures related to the technical features of the present invention, and do not strictly follow the actual scale to draw the complete structure and all details. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings.

[0056] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The phrase appearing in various locations in the specification does not necessarily refer to the same embodiment, nor is it limited to mutually exclusive independent or alternative embodiments. Those skilled in the art should be able to understand that the embodiments herein may be combined with other embodiments without causing structural conflicts.

[0057] In the description of this article, unless otherwise clearly specified and limited, the technical terms "installed", "connected", "connected" and the like should be understood in a broad sense, for example, it can be a movable connection, a fixed connection or an integral connection. For ordinary technicians in this field, the specific meanings of the above terms in the embodiments of this application can be understood according to specific circumstances.

[0058] In the description of this document, terms indicating orientation or positional relationships, such as "up", "down", "left", "right", "horizontal", "vertical", "height", "length", and "width", are intended to accurately describe the embodiments and simplify the description, but are not intended to limit the parts or structures involved to have a specific orientation, be installed or operate in a specific orientation, and should not be construed as a limitation on the embodiments in this document.

[0059] In the description of this article, the terms "first", "second" and the like are only used to distinguish different objects, and cannot be understood as indicating relative importance or limiting the quantity, specific order or primary and secondary relationship of the described technical features. In the description of this article, the meaning of "plurality" is at least two.

[0060] In metal bonding, it is usually necessary to monitor the bonding quality, monitor the welding head deformation, pressure, ultrasonic voltage, current, power and other signals, and then use these signals to determine the welding quality, such as cold welding, over welding, broken wire, etc. At the same time, record these data, save them locally or upload them to the server for subsequent tracking. The general method is to manually set the judgment range, either set the data interval of the entire bonding process, or set the data range of certain time points. If the setting is reasonable, it can also play a certain role in quality monitoring. If the setting is not good, it will be counterproductive, resulting in misjudgment, shutdown or missed judgment. The main problem of the prior art is that the threshold needs to be set manually. If the setting is too large, it will miss the judgment. If the setting is too small, it will be misjudged. Moreover, because the manual setting is relatively subjective, it is not conducive to the control of bonding quality. Manual settings cannot set too many judgment conditions. If there are too many judgment conditions, the difficulty of manually finding an accurate threshold will increase, reducing the accuracy of the threshold setting. If there are too few judgment conditions, the bonding quality cannot be accurately controlled.

[0061] In order to better achieve bonding quality control, an embodiment of one aspect of the present invention provides an ultrasonic bonding machine, which includes a bonding head and a quality detection system.

[0062] like Figure 1 and Figure 2 As shown, the bonding head includes an upper bracket 21, a lower bracket 22, a pressure motor 30, a working part 1 and a detection device. The pressure motor 30 can vibrate longitudinally, and the axial ends (i.e., the two ends in the vibration direction) of the pressure motor 30 are connected to the pressure spring sheet 31, and are connected to the upper bracket 21 and the lower bracket 22 through the pressure spring sheet 31. The working part 1 includes a cutter 11, a splitter 12 and a wire guide nozzle 13. The pressure motor 30 drives the working part 1 to vibrate through the transducer 24 to implement the bonding operation.

[0063] The detection device can detect the bonding parameters of the bonding head during operation, and the bonding parameters include a combination of at least two of ultrasonic current, ultrasonic voltage, ultrasonic power, and deformation. In a preferred embodiment, the detection device includes a grating ruler 41 fixedly connected to the pressure motor 30 and a reading head 42 fixedly connected to the lower bracket 22.

[0064] The structure of the pressure spring 31 is as follows: Figure 3 As shown, the pressure spring sheet 31 is a circular structure as a whole and is coaxially arranged with the pressure motor 30. The pressure spring sheet is fixed to the pressure motor 30, the upper bracket 21 or the lower bracket 22 in a centrally symmetrical manner. Specifically, in a preferred embodiment, the pressure spring sheet is evenly arranged with a plurality of centrally symmetrical fixing points in the axial direction, and the fixing points include an external fixing point 311 and an internal fixing point 312, wherein the external fixing point 311 is connected to the upper bracket 21 or the lower bracket 22, and the internal fixing point 312 is connected to the pressure motor. Specifically, in a preferred embodiment, Figure 3As shown, there are three external fixing points 311 and three internal fixing points 312, respectively, and they are evenly arranged in the circumferential direction. In a further preferred embodiment, three through grooves 313 are provided on the pressure spring plate 31, and the through grooves 313 axially penetrate the pressure spring plate 31 and are distributed in a spiral shape. Specifically, each through groove 313 extends between an internal fixing point 312 and an external fixing point 311, and the internal fixing point 312 and the external fixing point 311 are not the two internal fixing points 312 and the external fixing points 311 that are closest to each other.

[0065] The circular pressure spring piece 31 ensures that the splitting knife 11 can move up and down and deform during the bonding process. The purpose of designing the pressure spring piece 31 into a circular shape is to ensure that the force on all sides of the bonding head is uniform during the deformation process, thereby ensuring that the splitting knife 11 will not swing forward, backward, left or right, and the splitting knife 11 is always perpendicular to the bonding surface.

[0066] By finely designing the structure of the pressure spring 31, the grating ruler 41 can reciprocate along the axial direction of the pressure motor, and the deformation amount during the bonding process can be obtained at a frequency of 30M. The reading head 42 is fixed. Figure 4 As shown, the reading head 42 reads the displacement of the grating ruler 41 and outputs it as a TTL square wave signal. The upper computer motion control card obtains the deformation square wave signal output by the grating ruler at a speed of 4K / s. In the welding time of 200ms, 800 position data can be obtained, thereby realizing accurate reading of the bonding parameters. The resolution of the grating ruler can reach 0.1um, the sampling frequency can reach 30M, and the data is more accurate; the deformation is accurately matched with the timing of the ultrasonic signal to ensure the accuracy of the data. In addition to the deformation, the ultrasonic current, ultrasonic voltage, and ultrasonic power can be directly measured and obtained through the driving system of the ultrasonic bonding machine.

[0067] In a preferred embodiment, 60K and 80K ultrasonic transducers 24 are used, of which 60K is used for bonding aluminum wires and 80K is used for bonding copper wires. The ultrasonic power supply generates a 60K or 80K sinusoidal voltage to the piezoelectric ceramics. The piezoelectric ceramics are driven by the sinusoidal voltage to drive the amplitude transformer to vibrate back and forth, and the vibration amplitude is ≤±2.5μm.

[0068] The quality inspection system includes a memory and a processor. The memory can store the bonding parameter data corresponding to each operation in the bonding process. In a preferred embodiment, during the signal acquisition process, the synchronization of signal acquisition must be ensured, that is, the ultrasonic current, voltage signal and bonding deformation signal must start and end at the same time. This solution adopts the hardware DSP (central processing unit) to collect and cache in real time, and obtain it once after the bonding is completed to ensure data synchronization, interval consistency and real-time performance.

[0069] The memory also stores a quality detection program. When the processor executes the quality detection program, it can perform quality detection on each bonding operation according to the bonding parameters.

[0070] Specifically, the quality detection program uses a deep learning neural network model and is obtained in the following way:

[0071] Providing a bonding data set, wherein the bonding data set includes bonding samples for training, wherein the bonding samples are bonding parameter groups with labels, wherein the labels characterize the bonding quality, including normal, cold soldering, over soldering, and broken wire;

[0072] The bonding data set is divided into a training set and a test set. The deep learning neural network model is trained using the training set, and tested using the test set. When the accuracy of the test set label classification reaches the preset standard threshold, the training is completed and the deep learning neural network model is used as a quality detection program to evaluate the bonding quality.

[0073] In a preferred embodiment, a quality detection program is established based on the ResNet50 deep learning neural network.

[0074] For bonding with a wire diameter of 0.5mm, the dimensionless bonding power is 200, the bonding pressure is 1300g, the bonding time is 200ms, the search speed is 20000μm / s, the search height is 700μm, and the search force is 850.

[0075] The bond data set provided is shown in Table 1.

[0076] Table 1 Bonding data set

[0077]

[0078] The process of training ResNet50 is as follows Figure 5 shown.

[0079] Step 1: Use the third-party package pytorch in python to import the ResNet50 model, load the ResNet50 model, modify and expand it as needed, and add a fully connected layer for classification tasks. Use multiple convolution kernels to perform convolution operations to extract image features. Downsample the output of the convolution layer to reduce the size of the feature map. Send feature maps of different scales to different branches for processing to extract richer feature information. Map the processed feature vectors to categories for classification prediction. Normalize the output of the fully connected layer to obtain the probability distribution of four categories. Calculate the loss function based on the prediction results and the true labels to optimize the model parameters. Update the model weight parameters through the backpropagation algorithm to reduce the value of the loss function.

[0080] Step 2: Load the training weights. The initial weight parameters for training are the weights of the ResNet50 network used by the official author of ResNet50 on the ImageNet dataset. The advantage of this is that the weight parameters used here are the best data verified by multiple experiments and a large amount of data. By using the existing pre-trained model, the features and patterns learned on a large amount of data can be used to speed up the training of the model and improve the performance of the model on new tasks, making use of existing research results and saving time and cost. Transfer learning in machine learning is used here.

[0081] Step 3: Define input data and import labeled data sets, which include normal good products, cold solder joints, over solder joints, and broken wires, and are labeled 1234, where 1 is cold solder joints, 2 is over solder joints, 3 is broken wires, and 4 is normal good products. Image data preprocessing: data enhancement is performed on the acquired data, so that more data can be created based on limited data, thereby increasing the number and diversity of training samples, and thus improving the generalization ability and robustness of the model.

[0082] Step 4: forward propagation, such as Figure 6 As shown in the figure, the input data is passed to the input layer of the neural network. After the input layer, each neuron calculates the activation value based on the weight of its connection and the input data. The activation value of ResNet50 is calculated using the ReLU function. The activation value is passed to the next layer of neurons through the inter-layer connection of the neural network. This process is repeated in all layers of the neural network until it reaches the output layer. During this process, the neural network will calculate based on the input data and weights, and continuously adjust the weights to improve the prediction accuracy of new data.

[0083] The ReLU function is one of the common activation functions, and its expression is as follows:

[0084] f(x)=max(0,x),

[0085] The main advantage of the ReLU activation function is that it is computationally efficient and can avoid the problem of vanishing gradients when training deep neural networks. When the input is positive, the gradient of the ReLU activation function is 1, so backpropagation and parameter updates can be performed efficiently. When the input is negative, the gradient of the ReLU activation function is 0, so it will not affect the propagation of the gradient.

[0086] The ReLU activation function is a very important activation function in deep learning, and it has achieved good results in many applications.

[0087] The specific structure of the neural network includes: input layer: receiving input images; convolution layer 1: using 7x7 convolution kernel, stride 2, padding 3, to perform convolution operation, the purpose is to extract preliminary features of the image; maximum pooling layer 1: using 3x3 pooling kernel, stride 2, down-sample the output of convolution layer 1 to reduce the spatial size of the feature map; residual block 1: contains two convolution layers, the first convolution layer uses 3x3 convolution kernel, stride 1, padding 1, and the second convolution layer uses 1x1 convolution kernel. The role of these two convolutional layers is to transform the input features and perform residual learning; residual block 2: the structure is the same as residual block 1, but the number of convolution kernels is increased; residual block 3: similar to residual block 2, but the number of convolution kernels is further increased; residual block 4: this is the last residual block, with the same structure as residual block 3, but with more convolution kernels; global average pooling layer: perform global average pooling on the output of residual block 4 and map the features into a one-dimensional vector; fully connected layer: connect the one-dimensional vector to the classifier for classification prediction; output layer: output classification prediction results.

[0088] Among them, the role of global average pooling is:

[0089] 1. Reduce the number of parameters: Global average pooling converts the spatial dimension of the feature map into one dimension, thereby reducing the number of parameters of the model. This is very useful for processing high-resolution images or large datasets because it reduces computational costs and memory requirements.

[0090] 2. Translation invariance: Global average pooling performs global average on the input feature map and does not depend on the position information. This means that the model is invariant to the translation of the input image and will not produce different prediction results due to changes in the position of the image.

[0091] 3. Feature description: Global average pooling can be regarded as a feature description of the feature map. It averages the feature values ​​of each channel, reflecting the average characteristics of the channel over the entire image without considering the specific location information. This feature description can help the model capture the global information of the image and is very useful for some tasks that require classification or description of the entire image.

[0092] 4. Prevent overfitting: Global average pooling can prevent overfitting to a certain extent. Since it performs a global average on the input feature map, it reduces the model's dependence on local details, thereby reducing the complexity of the model and reducing the risk of overfitting.

[0093] Max Pooling divides the input image into several rectangular regions and outputs the maximum value for each sub-region. The functions of the max pooling layer are:

[0094] 1. Reduce dimensionality: Through pooling operations, the spatial size of the input feature map is reduced, thereby reducing the number of model parameters and the amount of computation, which helps to reduce overfitting.

[0095] 2. Feature selection: The pooling operation selects the maximum value in the input feature map and captures the most important feature information. This helps the model learn representative and discriminative features, which is very helpful for tasks such as image classification.

[0096] 3. Prevent overfitting: Because the pooling operation is insensitive to the spatial position of the input, it can prevent the model from being too dependent on the specific location information of the input, thereby reducing the risk of overfitting.

[0097] 4. Translation invariance: Maximum pooling has a certain degree of translation invariance. Even if the input image undergoes a small displacement in space, the result after pooling remains relatively stable, which makes the model robust to slight changes in the image.

[0098] 5. Increase generalization ability: Through the pooling operation, the model learns more general features, thereby enhancing its generalization ability for different input data.

[0099] The residual block consists of BTNK1 and BTNK2. The first residual block has two BTNK2 and one BTNK1. BTNK1 has four parameters: C, W, C1, S.

[0100] S: represents the step size in the convolution layer. When S is 1, the input size and output size are the same, indicating that no downsampling is performed.

[0101] C1: represents the number of feature maps output by the convolutional layer, that is, the number of output channels.

[0102] C: represents the number of input channels. The equality of C and C1 indicates that the 1×1 convolution layer on the left does not reduce the number of channels. In the last three stages, C=2*C1 indicates that the 1×1 convolution layer on the left reduces the number of channels.

[0103] W: represents the input size, namely length and width.

[0104] Compared with BTNK2, BTNK1 has inconsistent input and output channels. The right side of BINK1 first passes through a convolution layer to change the number of its output channels. The output is set to G(x). The G function matches the number of output channels on the left side. In this way, F and G are added and then passed through the Relu activation function to obtain the output of BINK1.

[0105] BTNK2 has two parameters: C and W.

[0106] C: represents the number of input channels. W: represents the input size.

[0107] The left side of BTNK2 passes through three convolution blocks (including BN and RELU), and its output is set to F(x). F and x are added and then passed through the Relu activation function to obtain the output of BINK2.

[0108] Step 5: Calculating loss generally refers to using the loss function to calculate the loss value. The loss value is used as a criterion for judging the quality of a model. The smaller the loss value, the better the effect of the model and the higher the accuracy. A normal model will continue to reduce the loss value through continuous training and eventually stabilize in a range. The loss function used in this article is the cross entropy loss function.

[0109] The cross entropy loss function is also known as logarithmic loss, log loss, or logistic loss. Each predicted class probability is compared to the actual class desired output of 0 or 1 and a score / loss is calculated, penalizing the probability based on how far it is from the actual expected value. The penalty is logarithmic in nature, producing large scores for large differences close to 1 and small scores for small differences tending towards 0.

[0110] Cross entropy loss is used when adjusting model weights during training. The goal is to minimize the loss, i.e. the smaller the loss, the better the model. A perfect model would have a cross entropy loss of 0.

[0111] Step 6: Backpropagation. Backpropagation is short for "error backpropagation". It is a common method used in conjunction with optimization methods (such as gradient descent) to train artificial neural networks. This method calculates the gradient of the loss function for all weights in the network. This gradient is fed back to the optimization method to update the weights to minimize the loss function. In ResNet50, the backpropagation algorithm calculates the gradient of the loss function for each convolutional layer, pooling layer, and fully connected layer based on the current input data and model parameters, such as Figure 7 These gradients are then back-propagated to the previous layer to update the parameters of the previous layer.

[0112] Step 7: Use the calculated gradient to adjust the weights according to the gradient descent algorithm or other optimization algorithms. ResNet50 usually uses the stochastic gradient descent (SGD) algorithm to update the weights. Stochastic gradient descent is a commonly used optimization algorithm used to find the optimal parameters of the model in deep learning. Specifically, when training ResNet50, the stochastic gradient descent algorithm randomly selects a portion of the training samples (called a mini-batch) and calculates the gradient of the model on these samples. It then adjusts the weights based on the direction and size of the gradient. In each iteration, the stochastic gradient descent algorithm updates the weights so that the value of the loss function decreases.

[0113] The core idea of ​​SGD is to randomly select a sample (or a small batch of samples) in each iteration to estimate the gradient, rather than using the entire dataset. The advantage of this is high computational efficiency, especially when the dataset is large. SGD is also able to escape from local minima because randomness introduces a certain amount of noise, which helps the model explore more parameter space.

[0114] The specific formula is as follows:

[0115] θ t+1 =θ t -η▽f i (θ t ).

[0116] Among them, θ represents the model parameters; t represents the current number of iterations; η is the learning rate, which controls the size of the step; f i is the loss function for the i-th data point or data batch. This formula means that in each iteration, we calculate the current parameter θ t The gradient f i (θ t ), and then update the parameters in the opposite direction of the gradient to reduce the value of the loss function.

[0117] Step 8: Use the test set to evaluate the model. For the classification problem in this article, multi-classification cross entropy is generally used for evaluation. Cross entropy is an evaluation indicator for classification problems, which reflects the difference between the predicted probability distribution and the true probability distribution. In a multi-classification task, each category has a probability distribution.

[0118] Step 9: Save the trained model to facilitate prediction of subsequent bonding quality classification.

[0119] To complete the entire model training process, during the operation, the collected data is divided into four categories, namely, cold soldering, over soldering, broken wire, and normal, and marked. When the classified label is less than 4, the product is considered unqualified, and when it is equal to 4, the product is considered qualified.

[0120] Finally, the trained model is tested using the test set, and deployed to the device for production operation after passing the test. In the preferred embodiment, if occasional misjudgments are found in the actual production process, the misjudged data can be directly marked to the corresponding category, manually clicked or the system automatically trained after reaching a certain number, and then the trained model is deployed to the device for operation. After the training is completed, the accuracy of the classification tested with the above test set is 95.6%, the inference time for each data is 8ms, and a multi-threaded method can be used to make a judgment when moving to the next welding point, which will not affect the UPH of the whole machine.

[0121] In a preferred embodiment, the process of establishing a quality inspection program is as follows Figure 7 The quality inspection program is used to inspect and classify the bonding parameters of a batch of products. The bonding power curve of the qualified products (normal state) obtained by inspection and identification is shown as follows: Figure 8a The deformation variable curve is shown as Figure 8b As shown; the bonding power curve of the broken wire state is as follows Figure 9a The deformation variable curve is shown as Figure 9b As shown; the bonding power curve of the over-welding state is as follows Fig.10a The deformation variable curve is shown as Fig.10b As shown; the bonding power curve of the cold solder joint state is as follows Fig.11a The deformation variable curve is shown as Fig.11b shown.

[0122] The ultrasonic bonding machine provided in the above embodiment uses a deep learning algorithm for quality control, which can effectively improve the efficiency of quality inspection, and can effectively improve the recognition accuracy by performing comprehensive inspection in multiple parameter dimensions. At the same time, the model can be updated and iterated using on-line production data according to the actual production scenario of the ultrasonic bonding machine, thereby improving the flexibility of the production line.

[0123] In a preferred embodiment, the memory also includes a cache device. During the signal acquisition process, the synchronization of signal acquisition must be ensured, that is, the ultrasonic current and voltage signals and the bonding deformation signal must start and end at the same time. The data can be synchronized to the processor by acquiring them once after the bonding is completed to ensure the consistency and real-time performance of the data interval.

[0124] The purpose of the above embodiments is to further explain the present invention in detail in conjunction with the accompanying drawings so that those skilled in the art can understand the technical concept of the present invention. Within the scope of the present invention, the optimization or equivalent replacement of the part structure or method steps involved, and the combination of implementation methods in different embodiments without conflict of structure and principle, all fall within the protection scope of the present invention.

Claims

1. An ultrasonic bonding machine, characterized in that: Includes bonding head and quality inspection system, among which, The bonding head includes a bracket, a working part, a pressure motor and a detection device. The working part, the pressure motor and the detection device are connected to the bracket. The pressure motor drives the working part to vibrate, The detection device detects the bonding parameters of the bonding head during operation, wherein the bonding parameters include at least two of ultrasonic current, ultrasonic voltage, ultrasonic power, and deformation amount; The quality detection system comprises a memory and a processor, wherein the memory stores a quality detection program and the bonding parameters detected by the detection device. The processor is capable of executing the quality detection program to detect the bonding quality of the bonding operation position corresponding to the bonding parameter, The quality inspection procedure is obtained by: Providing a bonding data set, the bonding data set comprising a bonding sample, the bonding sample comprising a bonding parameter group with a label, the label comprising normal, cold soldering, over soldering and broken wire; A deep learning neural network model is established, the bond data set is divided into a training set and a test set, the training set is used for training, and the test set is used for testing until the accuracy of label classification of the test set is no less than a given threshold.

2. The ultrasonic bonding machine according to claim 1, characterized in that: It also includes pressure springs, two of which are arranged at the two ends of the pressure motor in the vibration direction, so that the pressure motor is connected to the bracket through the pressure springs; the pressure springs are configured in a circular shape and are arranged coaxially with the pressure motor.

3. The ultrasonic bonding machine according to claim 2, characterized in that: The pressure spring sheet is provided with a plurality of internal fixing points and a plurality of external fixing points, wherein the internal fixing points and the external fixing points are respectively arranged in a centrally symmetrical manner, the internal fixing points are connected to the pressure motor, and the external fixing points are connected to the bracket.

4. The ultrasonic bonding machine according to claim 3, characterized in that: The pressure spring sheet is provided with a plurality of through grooves, which penetrate the pressure spring sheet axially and are centrally symmetrically arranged, and each of the through grooves extends between one of the outer fixing points and one of the inner fixing points.

5. The ultrasonic bonding machine according to claim 4, characterized in that: The through grooves are distributed in a spiral shape.

6. The ultrasonic bonding machine according to any one of claims 2 to 5, characterized in that: The detection device includes a matching grating and a reading head, wherein: The grating is fixedly arranged relative to the pressure motor, and the reading head is fixedly arranged relative to the bracket.

7. The ultrasonic bonding machine according to claim 1, characterized in that: The neural network model includes a ResNet50 deep learning neural network.

8. The ultrasonic bonding machine according to claim 7, characterized in that: The process of establishing a deep learning neural network is as follows: Use Python to import the ResNet50 model, add a fully connected layer for classification, normalize the output of the fully connected layer, and obtain the probability distribution of four categories: cold soldering, over soldering, broken wire, and normal. Calculate the loss function based on the predicted results and the true labels, and update the model weight parameters through the back propagation algorithm; Use transfer learning to load existing training weights; The labels are defined, values ​​are assigned to cold soldering, over soldering, broken wire and normal respectively, and data is input; The input data is passed to the input layer of the neural network, the activation value is calculated using the ReLU function, and it is passed to the next layer of neurons layer by layer through the inter-layer connections of the neural network until the output layer; Calculating a loss function, wherein the loss function includes a cross entropy loss function; Calculate the gradient of the loss function for each layer of neurons, and back-propagate to update the parameters of the neurons in the previous layer; Iterative calculation is performed to minimize the loss function and complete the training.

9. A bonding method, characterized in that: The bonding method uses the ultrasonic bonding machine as described in any one of claims 1 to 8 to perform bonding processing.

10. The bonding method according to claim 9, characterized in that: During the bonding process, the bonding data set is updated according to actual production data.

11. The bonding method according to claim 9 or 10, characterized in that: The memory includes a cache device, which stores the bonding parameters during the bonding process and transmits the bonding parameters to the processor after a given number of bonding processes.

Images