Wire bonding defect detection apparatus and method of operation thereof

Abstract

The present invention relates to an apparatus and method for detecting a joint defect occurring between a battery cell and a busbar connected via a wire by ultrasonic joint, in which machine learning using a convolutional neural network is used to learn data formed by collecting ultrasonic joint parameters over time, and then detect a joint defect based thereon.

Description

Technical Field

[0001] This application claims the benefit of priority to Korean Patent Application No. 2021-0112586, filed on August 25, 2021, the disclosure of which is incorporated herein by reference in its entirety.

[0002] This invention relates to an apparatus and method for detecting various defects occurring during a wire bonding process. More specifically, this invention relates to an apparatus and method for detecting defects occurring during an ultrasonic wire bonding process by sampling bonding parameters generated during the wire bonding process and machine learning using a convolutional neural network (CNN), a deep learning-based learning algorithm. Background Technology

[0003] Current wire bonding inspection methods, performed during the manufacturing of cylindrical battery packs with wire bonding, include manual visual inspection by physically pulling or picking up all wires and automated inspection methods using image processing.

[0004] These inspection methods suffer from the following problems: they become inefficient due to the increased size of the battery pack, and their reliability is difficult to ensure due to physical damage, camera position, resolution, and changes in the surrounding environment such as lighting.

[0005] Additionally, when separate equipment is used for additional electrical or mechanical inspections during the bonding process, other physical and electrical damage is involved, and additional costs and time are required during the inspection process, despite advantages in terms of wire defects, solder defects, wire height, and distance between adjacent wires.

[0006] If all possible bonding parameters that could affect quality during wire bonding are collected and analyzed in real time, it may be possible to perform highly reliable non-destructive inspections, thus accurately checking for errors in quality and process without physical damage.

[0007] Korean Patent Application Publication No. 2019-0081614 relates to a system for inspecting the weld quality of a weld using ultrasound, which discloses a technique that uses radiated ultrasound and echo ultrasound returned by machine learning as input. However, in this invention, the parameters in the joining step are not used as input, and the neural network is only used for machine learning.

[0008] Japanese Patent Application Publication No. 2019-185580 discloses a method for detecting anomalies occurring in a factory production facility using a convolutional neural network; however, this disclosure differs from the present invention in terms of substantial implementation or applicable objectives.

[0009] (Existing technical literature)

[0010] (Patent Document 1) Korean Patent Application Publication No. 2019-0081614

[0011] (Patent Document 2) Japanese Patent Application Publication No. 2019-185580 Summary of the Invention

[0012] Technical issues

[0013] In view of the above problems, the present invention has been proposed, and the object of the present invention is to provide an apparatus and method for detecting defects by sampling direct parameters that may affect wire bonding and non-destructive inspection using machine learning.

[0014] Technical solution

[0015] The wire joint defect detection apparatus according to the present invention for achieving the above-mentioned objectives comprises: an input data collection unit configured to continuously collect multiple ultrasonic joint parameters hourly and perform sampling; a data generation unit configured to apply wavelet transform to the collected data to convert the collected data into an image having a two-dimensional array; a convolutional neural network configured to perform machine learning on training data generated by the data generation unit; and a defect determination unit configured to use the trained convolutional neural network to determine defects caused by ultrasonic welding.

[0016] In another aspect of the implementation, the wire bonding defect detection method according to the present invention includes: an input data collection step, which continuously collects multiple ultrasonic bonding parameters hourly and performs sampling; a training data generation step, which converts the collected data into an image with a two-dimensional array by wavelet transform; a training step, which trains a convolutional neural network using the generated training data; and a defect determination step, which uses the convolutional neural network to determine defects caused by ultrasonic bonding.

[0017] Beneficial effects

[0018] As is evident from the above description, the advantage of the apparatus and method of using the apparatus according to the invention, which are capable of detecting defects caused by ultrasonic wire bonding, is that bonding parameters that may affect the quality of the wire bonding are extracted and the probability of defects is calculated through machine learning. Thus, efficiency and battery pack quality can be improved due to non-destructive inspection. Attached Figure Description

[0019] Figure 1 This is a view showing the basic structure of the battery cells and busbars connected to each other by wires.

[0020] Figure 2 This is a schematic view illustrating the construction of a wire bonding defect detection device according to an embodiment of the present invention.

[0021] Figure 3 This is a view illustrating CWT calculation using the PyTorch API and parameter values ​​that can be detected during the wire bonding process according to an embodiment of the present invention.

[0022] Figure 4 This is a view showing image data with a 50×39 array as a result of CWT calculation according to an embodiment of the present invention.

[0023] Figure 5 This is a view illustrating the basic structure of a convolutional neural network according to an embodiment of the present invention.

[0024] Figure 6 This is a view illustrating an embodiment of a convolutional neural network using the PyTorch API according to an embodiment of the present invention. Detailed Implementation

[0025] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that these preferred embodiments can be readily practiced by those skilled in the art to which the present invention pertains.

[0026] However, when describing the operating principles of the preferred embodiments of the present invention in detail, detailed descriptions of known functions and configurations incorporated herein may obscure the subject matter of the present invention, such descriptions will be omitted.

[0027] In addition, the same reference numerals will be used throughout the accompanying drawings to refer to parts that perform similar functions or operations.

[0028] When a component is referred to as being connected to another component throughout the specification, a component can be directly connected to another component, and a component can also be indirectly connected to another component via yet another component.

[0029] In addition, including an element does not mean excluding other elements, but rather that such elements may be included unless otherwise specified.

[0030] The invention will be described in more detail below.

[0031] Figure 1 This is a view showing the basic structure of the battery cell's electrodes 11 and busbar 12 connected to each other by wires, wherein the battery cell's electrodes 11 and busbar 12 are connected to each other by ultrasonic welding via aluminum wires 13.

[0032] During this process, most quality defects occur at joint 14, which is the soldered area to which the wires are connected.

[0033] Typically, wire bonding is a technique that uses metal wires to join two spaced-apart pads together. It mainly uses thermal compression and ultrasonic welding methods. In thermal compression, heat is applied for compression, while in ultrasonic welding, ultrasound is applied for bonding instead of heat.

[0034] The thermal compression method is a method in which the temperature of the bonding pads is raised to about 200°C in advance, the temperature of the end of the capillary is also raised to create the wire in a spherical shape, and the wire is attached while pressure is applied to the bonding pads through the capillary.

[0035] Ultrasonic welding is a method in which ultrasonic waves are applied to a wedge (a capillary-like wire-moving mechanism that does not form a ball) while the wire is lowered onto the pad to attach the wire to the pad. Its advantages include low cost in terms of process and materials. However, because ultrasonic waves are used instead of heat and pressure, a disadvantage of ultrasonic welding is that, while easy to handle, the tensile strength after bonding (the strength the wire can withstand when pulled after the connection) is low.

[0036] Specifically, since problems arising from the use of ultrasonic methods directly lead to product defects, it is especially necessary to accurately detect defects in real time in order to ensure product reliability.

[0037] In addition, the parameter values ​​required for ultrasonic injection may vary depending on the wire material (e.g., gold, aluminum, or copper), and the defect conditions may be set differently each time, which is troublesome.

[0038] The present invention is characterized in that, during the process of performing ultrasonic welding to electrically connect the electrodes and busbars of a battery cell to each other via aluminum wires, welding defects are determined by the electrical signal and frequency characteristic information applied to the joint point, which is the welding area.

[0039] Figure 2 This is a schematic view illustrating the construction of a wire bonding defect detection device, wherein the defect detection device 100 is characterized by including: a bonding parameter collection unit 110 configured to receive quality measurements from a device such as an ultrasonic wire bonding machine 200; a data generation unit 120 for machine learning; a convolutional neural network 130 configured to perform deep learning-based learning; and a defect determination unit 140 configured to determine whether a defect has occurred based on the learning results.

[0040] The junction parameter collection unit 110 is used to sequentially receive the current and voltage of the ultrasonic generator and the measured value of the conductor deformation based on the measurement time.

[0041] The data generation unit 120 samples the data received from the junction parameter collection unit 110 and uses a time / frequency transformation function to transform the input data for image recognition by the convolutional neural network 130.

[0042] The convolutional neural network 130 is a module that constitutes a deep learning algorithm for learning from two-dimensional array data as input.

[0043] The defect determination unit 140 is used to analyze the final result after applying test data to the trained convolutional neural network 130, thereby determining whether a defect has occurred, and can provide a "yes / no" result of determining whether a defect has occurred or perform classification into a group of specified defect types or forms; however, the invention is not limited thereto.

[0044] In addition, ultrasonic generator current, ultrasonic generator voltage, wire deformation (a measurement of the height of the aluminum wire as it melts), ultrasonic frequency, set voltage, and ultrasonic phase difference are also parameters that affect the wire bonding process.

[0045] In this invention, three parameters, namely ultrasonic generator current, ultrasonic generator voltage, and wire deformation, are set as combined parameters for mass measurement, and signal information such as phase information, ultrasonic frequency, resonant frequency, and pressure can be used as needed; however, the invention is not limited thereto.

[0046] Ultrasonic coupling parameters can be collected for a time period ranging from 10 ms to 1000 ms, preferably from 50 ms to 500 ms, and more preferably from 100 ms to 150 ms. If the time deviates from the above range, it is impossible to ensure effective parameters.

[0047] The time interval for collecting ultrasonic bonding parameters can be from 0.1 ms to 100 ms, preferably from 0.5 ms to 50 ms, and more preferably from 1 ms to 5 ms. If the collection time interval deviates from the above range, it is impossible to ensure effective parameters.

[0048] In one embodiment of the invention, 130 values, which are ultrasonic bonding parameters, are collected as input data over a period of 130 ms at 1 ms intervals. Preferably, 50 of these 130 values ​​are sampled. If more than 50 values ​​are sampled, there is a problem that there is no performance advantage, and the learning time increases exponentially.

[0049] In the sampling method, arbitrary selection is performed, or only data from the first 50ms is extracted, arranged, and converted to values ​​x from 0 to 1 using the following equation (which is a min-max scaling). new .

[0050]

[0051] Perform a continuous wavelet transform (CWT) to reconstruct the transformed value x in the time domain. new .

[0052] In this invention, CWT uses a Mexican Hat function as the waveform transform function to overlap with the original signal, generating 13 two-dimensional images while changing the scale value, one of the wavelet parameters, from 2 to 14. The scale value is not limited to 13, and various changes can be performed as needed. In addition to the Mexican Hat function, Bump, Morlet, or Paul functions can also be used as the waveform transform function.

[0053] Figure 3 This is a view showing the engagement parameters 300 received from the engagement parameter collection unit 110 according to an embodiment of the present invention. Figure 3 The following briefly illustrates an implementation method 310 that essentially uses the PyTorch API to perform CWT calculations.

[0054] Figure 4 This is a view showing image data 400 with a 50×39 two-dimensional array as the result of CWT calculation according to an embodiment of the present invention, wherein the image data 400 with the two-dimensional array is applied to a convolutional neural network as basic training data for machine learning.

[0055] In this invention, the analysis performed by CWT is an analysis method mainly used for fault diagnosis. Its advantage lies in discovering the discontinuities of signals that cannot be seen in Fourier transform, and it is beneficial for analyzing pulse signals with large amplitude in a very short time.

[0056] The neural network-based learning algorithm used in this invention employs a deep learning model with a multi-layered neural network structure, including an input layer, a hidden layer, and an output layer, and primarily uses a convolutional neural network, which is used when processing spatial shape information of image data.

[0057] The basic model of a general CNN is characterized by including: convolutional layers, which are configured to extract local features by using filters to compute various convolutions on the image; pooling layers, which are configured to perform downsampling to reduce data size while preserving spatial information; and fully connected multilayer neural networks, which are used for final classification.

[0058] Compared to general multilayer neural networks, the convolutional neural network in this invention can minimize the size of the learning parameters. Therefore, the scale of fully connected multilayer neural networks can be reduced, allowing for the design of structures that can learn in a short time. Furthermore, features can be extracted from ultrasonic welding more effectively than direct numerical methods.

[0059] Figure 5 This is a view showing the structure of the convolutional neural network implemented in this invention.

[0060] As from Figure 5 As can be seen, convolutional neural networks typically consist of convolutional layers, GAP layers, and fully connected layers (also known as affine layers), and convolutional layers can be designed to have four layers.

[0061] The convolutional layer consists of four layers, and different numbers of filters, namely 32, 64, 128 and 256 filters, are applied to the corresponding layers.

[0062] In the convolutional layer, a 3×3 two-dimensional array of filters is used, and the stride for the filter movement interval is set to 1. Additionally, the padding value required to resize the output is also set to 1 so that the output has the same size as the input image.

[0063] In addition, the value of each unit is changed by the following equation, Leaky ReLU (rectified linear unit), which is used as the activation function.

[0064]

[0065] The image size is halved by pooling calculations after the activation function described above is used to modify the image.

[0066] In this invention, the Max Pooling technique is used for pooling calculations. This Max Pooling technique extracts the maximum value in a region set to a 2×2 filter size as the representative value.

[0067] In addition to 2×2, 3×3 can be applied differently for the filter size of each level.

[0068] The first layer of the convolutional layer is the layer in the first step that receives training data (images) with an array of 39×50. It generates 32 layers of convolutional images using 32 filters and performs activation functions and pooling calculations, thereby finally completing a 32-layer image with an array of 20×25.

[0069] When the above process is repeated, a 64-layer image with a size of 10×13 can be constructed in the second level, a 128-layer image with a size of 5×7 can be constructed in the third level, and a 256-layer image with a size of 3×4 can be constructed in the fourth level, which is the last level.

[0070] The image generated in the fourth level can be not directly connected to the fully connected level. Instead, a separate global average polling (GAP) level can be set in between to reduce the 3×4×256 image (feature information) to a size of 256. This allows for lossless processing of the feature information included in the filter while reducing the number of computations.

[0071] GAP can be used to average all the values ​​at each level or simply to sum them; however, the invention is not limited thereto.

[0072] The result of GAP, which is constructed as represented by the following equation, is constructed as having k one-dimensional arrays and is transmitted to the fully connected layer, where k is 256.

[0073]

[0074] Figure 6 This is a view illustrating an implementation of a convolutional neural network algorithm using the PyTorch API, where convolutional computations via the conv2d, LeakyReLU, and MaxPool2d APIs are repeatedly performed over a specified number of layers, and linear APIs are executed for computations in the AvgPool2d and fully connected layers based on the results of the convolutional computations. The parameter values ​​in each API are not restricted and can be varied as needed within the skill level of a person skilled in the art.

[0075] The conv2d API allows a windowed filter with weight parameter values ​​to be multiplied by the input data while moving the filter over a predetermined interval, and can perform calculations to obtain a sum.

[0076] The LeakyReLU API (which sets the activation function) can set the negative_slope value to 0.01, thereby preventing negative input values ​​from being deactivated to 0 and maintaining a small slope with a constant value, thus enabling effective learning.

[0077] The MaxPool2d API (which is a function that extracts the maximum value in a specified region as a representative value) is used to construct a filter with a size of 2×2, thereby reducing the image size by half to reduce the number of computations in the fully connected layers. The stride is set to 2, the padding is set to 1, and the ceil_mode is set to FALSE, thus constructing the filter with integers.

[0078] The AvgPool2d API is used as... Figure 5 The fourth level of convolution calculation in the image transforms the 256 layers of a 3×4 image into 256 one-dimensional data.

[0079] The linear API, which is a function that performs learning through weight computation in a fully connected layer, can take 256 input values ​​and 4 classification result values ​​as parameters.

[0080] Although the specific details of the invention have been described in detail, those skilled in the art will understand that the detailed description only discloses preferred embodiments of the invention and therefore does not limit the scope of the invention. Consequently, those skilled in the art will understand that various changes and modifications are possible without departing from the category and concept of the invention, and it will be apparent that such changes and modifications fall within the scope of the appended claims.

[0081] (Description of reference numerals in the attached diagram)

[0082] 11: Electrode

[0083] 12: Busbar

[0084] 13: Wire

[0085] 14: Joint point

[0086] 100: Defect Detection Device

[0087] 110: Connection parameter collection unit

[0088] 120: Data Generation Unit

[0089] 130: Convolutional Neural Networks

[0090] 140: Defect Determination Unit

[0091] 200: Ultrasonic wire bonding machine

[0092] 300: Connection parameter

[0093] 400: Training image data

Claims

1. An apparatus for detecting a bonding defect occurring between a battery cell and a busbar, the battery cell and the busbar being connected to each other via wires by ultrasonic bonding, the apparatus comprising: An input data collection unit is configured to continuously collect multiple ultrasonic coupling parameters over time and perform sampling. A data generation unit configured to apply wavelet transform to the collected data in order to convert the collected data into an image with a two-dimensional array; A convolutional neural network configured to perform machine learning on training data generated by the data generation unit; as well as A defect determination unit is configured to use a trained convolutional neural network to determine defects caused by joining. The ultrasonic bonding parameters include the measured value of conductor deformation, and The measured value of the conductor deformation includes a measured value of the height of the conductor used to connect the battery cell and the busbar, which changes as the conductor melts.

2. The apparatus according to claim 1, wherein, The ultrasonic bonding parameters also include ultrasonic generator current and ultrasonic generator voltage.

3. The apparatus according to claim 1, wherein, The wavelet transform converts the input data to values ​​from 0 to 1 using a minimum-maximum scaling method, and generates training data with the two-dimensional array using a Mexican hat waveform transform function with a scale value of 2 or greater.

4. The apparatus according to claim 1, wherein, The convolutional neural network consists of four convolutional layers, a global average polling (GAP) layer, and a fully connected layer.

5. The apparatus according to claim 1, wherein, The ultrasonic bonding parameters were collected over a period of 10 ms to 1000 ms.

6. The apparatus according to claim 1, wherein, The ultrasonic bonding parameters were collected at intervals ranging from 0.1 ms to 100 ms.

7. A method for detecting a bonding defect occurring between a battery cell and a busbar, the battery cell and the busbar being connected to each other via wires by ultrasonic bonding, the method comprising the steps of: (a) An input data collection step, wherein the input data collection step continuously collects multiple ultrasonic coupling parameters over time and performs sampling; (b) A training data generation step, wherein the training data generation step converts the collected data into an image with a two-dimensional array through wavelet transform; (c) The steps of training the convolutional neural network using the generated training data; and (d) A defect determination step, wherein the defect determination step uses the convolutional neural network to determine defects caused by the ultrasonic bonding. The ultrasonic bonding parameters include the measured value of conductor deformation, and The measured value of the conductor deformation includes a measured value of the height of the conductor used to connect the battery cell and the busbar, which changes as the conductor melts.

8. The method according to claim 7, wherein, The wavelet transform in step (b) is configured to convert the input data to values ​​from 0 to 1 by min-max scaling and to generate the wavelet transform of the training data with the two-dimensional array using a scale value of 2 or greater using the Mexican hat waveform transform function.

9. The method according to claim 7, wherein, Step (c) performs convolution computation on the image with the two-dimensional array generated in step (b) using a 2 × 2 or 3 × 3 filter, and performs computation by activation function and pooling to generate 256 one-dimensional data.

10. The method according to claim 7 or 8, wherein, The ultrasonic bonding parameters include ultrasonic generator current and ultrasonic generator voltage.

Images