Bonding machine welding quality detection method based on machine learning

Abstract

The invention discloses a bonding machine welding quality detection method based on machine learning. The method comprises the following steps that deformation data and ultrasonic impedance data of welding spots are continuously collected in the welding process of a bonding machine; when the number of the currently collected welding spots reaches a set value, feature processing is carried out on the average deformation rate and the minimum impedance of the currently collected welding spots, and clustering is carried out based on feature data of feature processing; continuously collecting the welding spots and judging whether the new welding spots are normally welded or not based on the clustered boundaries; if the welding of the new welding spot is normal, adding the characteristics of the new welding spot into a cluster for cluster updating; if the welding of the new welding spot is abnormal, alarming and interrupting the welding process; and the steps are repeated until all the welding processes are completed. The method has the beneficial effects that compared with an electric detection method, the detection method has the advantages that the welding result is judged through the transducer, the deformation amount and other existing characteristics, no additional device is needed, cost is saved, and the situation that welding can be conducted but the welding quality is poor can be distinguished.

Description

Technical Field

[0001] The present application relates to the technical field of bonding machines, and in particular to a bonding machine welding quality detection method based on machine learning. Background Art

[0002] During the bonding process of the bonding machine, welding quality problems caused by chip / substrate contamination are often difficult to detect directly during the welding process. Even if welding quality problems can be detected in subsequent processes, it will lead to missed opportunities for re-welding, resulting in chip scrapping and reduced yield. If the quality problems of the welding wire cannot be detected in time, it may even cause subsequent welding points to be empty (referring to direct contact between the splitter and the chip / substrate), causing damage to the chip / substrate.

[0003] At present, the main methods for welding quality inspection include electrical inspection method, deformation detection, etc. Among them, the electrical inspection method uses the difference in current / voltage when the circuit is conductive / non-conductive to detect the connectivity between the welding wire and the chip / substrate. It has better detection capabilities for welding points that cannot be welded at all or welding wire breakage, and can also better detect the length of the wire tail.

[0004] However, the electrical inspection method cannot effectively detect welds that can be welded but do not meet the welding quality standards. In addition, the electrical inspection method requires a special power-on circuit, which increases the complexity and cost of the bonding head. At the same time, the electrical inspection method often requires additional time or action for detection, which will slow down the wire bonding speed. Summary of the invention

[0005] One of the objectives of the present application is to provide a bonding machine welding quality detection method based on machine learning that can solve at least one of the defects in the above-mentioned background technology.

[0006] In order to achieve at least one of the above purposes, the technical solution adopted in the present application is: a bonding machine welding quality detection method based on machine learning, comprising the following steps:

[0007] S100: Continuously collect deformation data and ultrasonic impedance data of solder joints during the bonding process;

[0008] S200: When the number of currently collected solder joints reaches a set value, feature processing is performed on the average deformation rate and the minimum impedance of the currently collected solder joints, and clustering is performed based on the feature data processed by the feature processing;

[0009] S300: Continue to collect welding points and determine whether the welding of new welding points is normal based on the boundaries of the cluster; if the welding of the new welding points is normal, add the characteristics of the new welding points to the cluster for cluster update; if the welding of the new welding points is abnormal, issue an alarm and interrupt the welding process;

[0010] S400: Repeat step S300 until the entire welding process is completed.

[0011] Preferably, in step S100, the deformation data and ultrasonic impedance data of each weld are sampled multiple times during the welding process of each weld, so that the average deformation rate and minimum impedance of the current weld are calculated and recorded after the welding of the weld is completed.

[0012] Preferably, the clustering adopts the optimized DBSCAN algorithm; in the process of clustering the solder joint feature data, the clustering input is adjusted in real time by observing the clustering result to update the clustering.

[0013] Preferably, the input quantity of clustering includes a domain radius and a threshold, and the clustering is updated by adjusting the threshold.

[0014] Preferably, in step S300, the welding condition determination of the new welding point includes the following specific processes:

[0015] S310: establishing a cluster coordinate system, where the horizontal and vertical axes represent the average deformation rate and the minimum impedance respectively;

[0016] S320: forming corresponding cluster points in a cluster coordinate system based on the clustering result;

[0017] S330: Find the core point in the cluster points, and divide the boundary according to the core point;

[0018] S340: Mapping the features corresponding to the new weld point to the cluster coordinate system to form a mapping point. If the mapping point is between the boundary and the horizontal and vertical coordinate axes, it is determined that the welding of the new weld point is abnormal; otherwise, it is determined that the welding of the new weld point is normal.

[0019] Preferably, in step S300, when the number of clustered welds reaches a set data update threshold, some historical welds are discarded.

[0020] Preferably, in step S300, when the number of collected welds reaches a set data update threshold, a set number of latest welds are retained and the retained welds are re-clustered, and the remaining welds are discarded.

[0021] Preferably, in step S300, if false alarms occur frequently or the actual welding quality is poor but no alarm is given, the input quantity of the cluster is adjusted based on the distribution of cluster points in the cluster coordinate system.

[0022] Compared with the prior art, the beneficial effects of this application are:

[0023] (1) Compared with the electrical inspection method, the inspection method of the present application utilizes existing characteristics such as transducers and deformation to judge the welding results, does not require additional equipment, saves costs, and can distinguish between situations where welding is possible but the welding quality is poor.

[0024] (2) Compared with the pure deformation detection method, the present application can be applied to the detection after the deformation control is turned on. The present application detects the deformation rate and the data related to the transducer as characteristic variables at the same time, and the detection result is more reliable.

[0025] (3) This application adopts an unsupervised learning method, which does not require the customer to input the welding quality evaluation results, is easy to operate, and uses a visual display to facilitate users to find a suitable threshold.

[0026] (4) This application uses an optimized DBSCAN clustering algorithm, which ignores the impact of new values ​​on historical values ​​and is optimized for the binary classification used in this application, with a fast calculation speed. Compared with the electrical inspection method, the additional time consumed by the detection method of this application is within 1ms, which has almost no impact on the speed of the wire bonding. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of deformation and impedance curves of the bonding machine when welding on normal and abnormal surfaces.

[0028] Figure 2 Schematic diagram of deformation time and minimum impedance of the bonding machine when welding on normal surface and abnormal surface.

[0029] Figure 3 The figure is a schematic diagram of the overall workflow of this application.

[0030] Figure 4 A clustering diagram of a specific input quantity of one solder joint and two solder joints in this application.

[0031] Figure 5 Based on this application Figure 4 The clustering diagram shown is after the input quantity of solder joint one and solder joint two is changed. DETAILED DESCRIPTION

[0032] Below, the present application is further described in conjunction with specific implementation methods. It should be noted that, under the premise of no conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0033] In the description of the present application, it should be noted that directional words, such as the terms "center", "lateral", "longitudinal", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc., indicating directions and positional relationships are based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of narrating the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and cannot be understood as limiting the specific scope of protection of the present application.

[0034] It should be noted that the terms "first", "second", etc. in the description and claims of the present application are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0035] The terms "including" and "having" and any variations thereof in the specification and claims of the present application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or apparatuses.

[0036] In order to facilitate the understanding of the following contents of this solution, we can first understand the impact of surface contaminated chips / substrates on welding quality, and then obtain the discrimination conditions for distinguishing abnormal welding on contaminated surfaces from welding on normal surfaces. This can be specifically demonstrated through welding comparison experiments on contaminated surfaces and normal surfaces.

[0037] The experiment used the fully automatic wedge bonder No. ZH6000-W#2004006 of Ningbo Shangjin Automation Co., Ltd. The wedge type was 4WDD4-2025-W5C-M02, MPP, and the welding wire was ordinary 38μm aluminum wire. The ambient temperature during welding was 20℃. The normal welding surface used a new gold-plated substrate, and the abnormal welding surface used a new gold-plated substrate contaminated by oil. In the experiment, the deformation, ultrasonic impedance, ultrasonic frequency, and ultrasonic current of the solder joint during the welding process were collected respectively. The sampling frequency was 1600Hz, the deformation measurement error was ±5e-4mm, and the impedance measurement accuracy was 1 / 65536Ω.

[0038] Based on the experimental results, the data was analyzed and found that in the constant current mode, the ultrasonic current and ultrasonic frequency of the normal surface and the abnormal surface during welding were constant and had no difference. Therefore, the main focus was on the deformation and ultrasonic impedance.

[0039] like Figure 1The figure shows the curves of aluminum wire diameter and ultrasonic impedance changing with welding time in a single experiment. Since the deformation control is turned on, the final deformation is consistent, and the deformation rate of the normal surface is higher than that of the oily surface; at the same time, the ultrasonic impedance of the normal surface during welding is significantly higher than that of the abnormal surface.

[0040] like Figure 2 As shown in FIG. 1 , a schematic diagram of minimum impedance and deformation time for normal surface welding and abnormal surface welding is shown. The minimum impedance is displayed in the form of data points on a two-dimensional graph, and the minimum impedance data points corresponding to the normal surface and the abnormal surface can form two easily distinguishable category areas on the plane.

[0041] In summary, from the above experimental results, it can be seen that the parameters that can characterize the welding quality are the ultrasonic impedance during the welding process, that is, the transducer impedance, and the deformation speed during the welding process; and the data points for different welding surfaces can be distinguished by themselves. Therefore, in this application, the ultrasonic impedance and deformation speed during the welding process can be used as characteristic values ​​for the detection of the bonding quality of the bonding machine, and the typical unsupervised learning method in machine learning - clustering is used to evaluate the welding quality, which will be described in detail below.

[0042] One of the preferred embodiments of the present application is as follows: Figure 1 As shown, a bonding machine welding quality detection method based on machine learning includes the following steps:

[0043] S100: Continuously collect deformation data and ultrasonic impedance data of solder joints during the bonding process.

[0044] S200: When the number of currently collected solder joints reaches a set value, feature processing is performed on the average deformation rate and the minimum impedance of the currently collected solder joints, and clustering is performed based on the feature data of the feature processing.

[0045] S300: Continue to collect welding points and determine whether the new welding points are welded normally based on the boundaries of the cluster; if the new welding points are welded normally, add the features of the new welding points to the cluster for cluster update; if the new welding points are welded abnormally, issue an alarm and interrupt the welding process.

[0046] S400: Repeat step S300 until the entire welding process is completed.

[0047] It is understandable that the core of this application is to collect the characteristic data of a small number of current welds, namely the deformation and ultrasonic impedance data, and then process and cluster the collected characteristic data. According to the aforementioned experimental content, the characteristic data of the welds will be divided into regions according to the welding quality of the welds, so the location of the clustering area that meets the welding quality can be found and the boundaries can be set. When new welds are collected later, the new welds can be compared with the current clustering boundaries for judgment. If the new welds meet the boundary conditions, it can be determined that the welding quality is qualified and integrated into the clustering update. If the new welds do not meet the boundary conditions, it is determined that the welding quality is unqualified, and an alarm needs to be issued and the welding process interrupted to avoid producing defective products.

[0048] The detection method of the present application can use existing characteristics such as transducers (ultrasonic impedance) and deformation to judge welding results, without the need for additional devices, which can effectively save costs and distinguish between situations where welding is possible but the welding quality is poor. And compared with the traditional electrical detection method, the present application can be applied to detection after the deformation control is turned on, making the detection results more reliable.

[0049] It should be known that the solder joints formed by the bonding machine are different according to the welding targets. The solder joints formed during the bonding welding process can be divided into one solder joint and two solder joints. Among them, one solder joint can also be called a pin solder joint, which refers to the solder joint that connects the pin of the chip to the external circuit; the second solder joint refers to the solder joint that connects the hook hole inside the chip to the external circuit. In the packaging welding process of the chip product, one solder joint and two solder joints will be generated at the same time, so the welding quality detection technology of this application can be used to perform quality analysis on both the first solder joint and the second solder joint.

[0050] In this embodiment, in step S100, deformation data and ultrasonic impedance data of each weld are collected multiple times during the welding process of each weld, so that the average deformation rate and minimum impedance of the current weld are calculated and recorded after the welding of the weld is completed, so that when performing step S200, the calculated average deformation rate and minimum impedance can be directly feature processed.

[0051] It is understandable that the number of data sampling times for a single welding point during the welding process can be selected according to the actual needs of technical personnel in this field. For example, the deformation amount and minimum impedance data can be sampled ten times during one welding process, which can minimize the measurement error.

[0052] Since the deformation of the weld can be regarded as a value that changes relatively evenly over time, it can be characterized by the average deformation rate of the weld after welding is completed. The average deformation rate can be calculated by first calculating the deformation rate corresponding to each adjacent sampling moment, and then accumulating the calculated deformation rate and dividing it by the total number of sampling intervals to obtain the average deformation rate; of course, the average deformation rate can also be obtained by weight distribution; the specific calculation process is a well-known technology for those skilled in the art, so it will not be elaborated in detail here.

[0053] The minimum impedance characterizes the welding quality of the solder joint, that is, the welding quality of the solder joint depends on whether the value of the minimum impedance meets the design requirements. For the calculation of the minimum impedance, when the ultrasonic impedance is sampled for the first time, the sampled ultrasonic impedance is recorded as the minimum impedance at the current moment; in the subsequent sampling process, at the current sampling moment, the sampled ultrasonic impedance is compared with the minimum impedance recorded at the previous moment. If the currently sampled ultrasonic impedance is less than the minimum impedance already recorded, the currently sampled ultrasonic impedance is recorded as the minimum impedance and overwrites the invalid minimum impedance. If the currently sampled ultrasonic impedance is greater than or equal to the minimum impedance already recorded, the minimum impedance already recorded continues to be stored.

[0054] In this embodiment, when performing step S200, there are multiple ways to process the characteristics of the average deformation rate and the minimum impedance of the solder joint, and a common way is normalization processing. Normalization processing can remove abnormal points in the characteristic parameters of the current solder joint, and map the characteristic parameters to a range of 0 to 1 to facilitate subsequent digital signal processing.

[0055] It should be known that the clustering performed in step S200 is the first clustering, and the set value of the number of solder joints required for the first clustering can be determined by the technicians in this field according to their actual needs. The set value should not be too large or too small. A too large value will make the judgment of some solder joints rough, and a too small value will affect the boundary accuracy. For example, the set value can be 10, 20, 50, etc.

[0056] In this embodiment, there are many specific implementation methods of clustering, that is, there are many types of clustering algorithms; such as partition-based clustering methods, hierarchical-based clustering methods, density-based clustering methods, and network-based clustering methods. The specific clustering method can be selected according to the actual needs of those skilled in the art. In one specific example, the optimized DBSCAN algorithm is preferably used for clustering; the DBSCAN algorithm is a density-based clustering algorithm that can ignore the impact of new values ​​on historical values, and the present application optimizes the DBSCAN algorithm for the binary classification used, so that the optimized DBSCAN algorithm has a faster calculation speed. Compared with the traditional electrical inspection method, the detection method of the present application consumes an additional time within 1ms, which hardly affects the welding speed of the welding wire. In the process of clustering the weld feature data, the input amount of the clustering can be adjusted in real time by observing the clustering results to update the clustering.

[0057] It should be known that the input of the DBSCAN algorithm includes the domain radius when defining density and the threshold when defining core points. When the clustering result does not meet expectations, the clustering result can be adjusted by changing the domain radius, or by changing the threshold. Taking into account the usage scenario of this embodiment, in this embodiment, the clustering can be updated by changing the threshold. It should be noted that after each update of the threshold, the characteristic parameters of the solder joints currently participating in the clustering need to be re-normalized before clustering.

[0058] In this embodiment, when performing step S300, the welding condition determination of the new welding point includes the following specific processes:

[0059] S310: Establish a cluster coordinate system, where the horizontal and vertical axes represent the average deformation rate and the minimum impedance respectively.

[0060] S320: forming corresponding cluster points in a cluster coordinate system based on the clustering result.

[0061] S330: Find the core point in the cluster points, and divide the boundary according to the core point.

[0062] S340: Mapping the features corresponding to the new weld point to the cluster coordinate system to form a mapping point. If the mapping point is between the boundary and the horizontal and vertical coordinate axes, it is determined that the welding of the new weld point is abnormal; otherwise, it is determined that the welding of the new weld point is normal.

[0063] To facilitate understanding, a detailed description will be given below using specific examples.

[0064] Specifically, Figure 4As shown in the figure, it is a schematic diagram of the clustering coordinate system formed when the thresholds input for welding quality inspection of solder joint 1 and solder joint 2 are both 0.2. The green points in the figure represent the core points of the optimized DBSCAN algorithm, and the blue points represent the boundary points of the optimized DBSCAN algorithm; then after finding the core points, the boundary points can be used as the boundaries of the clustering area. Since the red points are outside the boundaries of the clustering area, the red points represent points with abnormal welding.

[0065] Depend on Figure 4 From the distribution of cluster points, we can see that in the cluster coordinate system corresponding to the second weld point (the coordinate system below), there are only core points and abnormal welding points, but no boundary points. This will result in no judgment boundary when judging the welding quality of new weld points in the future, so the threshold of the second weld point needs to be adjusted. Figure 5 As shown in FIG. 1 , it is a schematic diagram of the clustering coordinate system of two welding points after the threshold value input based on welding quality detection is adjusted to 0.1. Figure 5 It can be seen that after the threshold adjustment is completed, boundary points appear in the cluster coordinate system corresponding to the two welding points.

[0066] In this embodiment, when performing step S300, when the number of clustered welds reaches the set data update threshold, some historical welds can be discarded. Specifically, since each time the characteristic parameters of a new weld are added, all the characteristic parameters of all current welds need to be re-normalized and then clustered and updated, as the number of welds gradually increases, the amount of data clustered by the algorithm will also gradually increase, which will lead to a decrease in the data processing speed of the algorithm. In addition, during a long welding process, the characteristics of the transducer itself may change, and too old data has little reference value, but may affect the accuracy of the clustering results and judgments. Therefore, when the amount of data reaches a certain level, it is necessary to discard some historical data, which can greatly improve the data processing efficiency of the algorithm and increase the data processing capacity of the algorithm exponentially.

[0067] It should be known that the specific setting of the data update threshold can be determined according to the actual needs of technical personnel in this field; for example, the value of the data update threshold can be 10,000. When the total number of welds to be clustered exceeds 10,000, some historical welds can be discarded to reduce the number of currently remaining welds, thereby effectively improving the data processing efficiency of the algorithm.

[0068] In this embodiment, when performing step S300, when the number of collected welds reaches the set data update threshold, the set number of latest welds is retained and the retained welds are re-clustered, and the remaining welds are discarded. Specifically, as the number of welds increases, when the threshold input to the DBSCAN algorithm remains unchanged, the core points will gradually diffuse, that is, the core points diverge as the number of clusters increases, which will cause the position of the boundary points to gradually approach the coordinate axis, thereby reducing the algorithm's detection accuracy for welding quality. Therefore, when the number of welds reaches the set data update threshold, it is necessary to discard the feature data of welds that are too old, and re-cluster the feature parameters of the remaining newer welds. And by reducing the scale of clustered data, the speed of clustering can be further improved to ensure that the entire welding quality judgment process takes less than 1ms to avoid affecting the welding speed.

[0069] It should be known that the number of welds to be retained can be selected according to the actual needs of the technicians in this field; for example, the number of welds to be retained is set to 5000, and the above data update threshold is taken as 10000 as an example. When the number of welds accumulates to 10000, the latest 5000 welds can be retained, and the 5000 historical welds recorded since the beginning of welding can be discarded.

[0070] In this embodiment, when performing step S300, if frequent false alarms occur or the actual welding quality is poor but no alarm is given, the input quantity of the cluster, such as the threshold, is adjusted based on the distribution of the cluster points in the cluster coordinate system. The adjustment of the threshold can take effect immediately and be reflected in the cluster coordinate system.

[0071] The above describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and the specification only describe the principles of the present application. The present application may have various changes and improvements without departing from the spirit and scope of the present application, and these changes and improvements fall within the scope of the present application for which protection is sought. The scope of protection claimed by the present application is defined by the attached claims and their equivalents.

Claims

1. A bonding machine welding quality detection method based on machine learning, characterized in that: The steps include: S100: Continuously collect deformation data and ultrasonic impedance data of solder joints during the bonding process; S200: When the number of currently collected solder joints reaches a set value, feature processing is performed on the average deformation rate and minimum impedance of the currently collected solder joints, and clustering is performed based on the feature data processed by the feature processing; S300: Continue to collect welding points and determine whether the welding of new welding points is normal based on the boundaries of the cluster; if the welding of new welding points is normal, add the characteristics of the new welding points to the cluster for cluster update; if the welding of new welding points is abnormal, issue an alarm and interrupt the welding process; S400: Repeat step S300 until the entire welding process is completed.

2. The bonding machine welding quality detection method based on machine learning as claimed in claim 1, characterized in that: In step S100, during the welding process of each welding point, the deformation data and ultrasonic impedance data of the welding point are sampled multiple times, so that after the welding of the welding point is completed, the average deformation rate and the minimum impedance of the current welding point are calculated and recorded.

3. The bonding machine welding quality detection method based on machine learning as claimed in claim 1, characterized in that: Clustering uses the optimized DBSCAN algorithm; In the process of clustering the solder joint feature data, the clustering input is adjusted in real time by observing the clustering results to update the clustering.

4. The bonding machine welding quality detection method based on machine learning as claimed in claim 3, characterized in that: The input of clustering includes the domain radius and threshold, and the clustering is updated by adjusting the threshold.

5. The bonding machine welding quality detection method based on machine learning according to any one of claims 1 to 4, characterized in that: In step S300, the welding condition determination of the new welding point includes the following specific processes: S310: establishing a cluster coordinate system, where the horizontal and vertical coordinates represent the average deformation rate and the minimum impedance respectively; S320: forming corresponding cluster points in a cluster coordinate system based on the clustering result; S330: Select cluster points close to the horizontal and vertical axes as boundaries; S340: Mapping the features corresponding to the new weld point to the cluster coordinate system to form a mapping point. If the mapping point is between the boundary and the horizontal and vertical coordinate axes, it is determined that the welding of the new weld point is abnormal; otherwise, it is determined that the welding of the new weld point is normal.

6. The bonding machine welding quality detection method based on machine learning as claimed in claim 5, characterized in that: In step S300, when the number of clustered welds reaches a set data update threshold, some historical welds are discarded.

7. The bonding machine welding quality detection method based on machine learning as claimed in claim 6, characterized in that: In step S300, when the number of collected welding points reaches a set data update threshold, a set number of latest welding points are retained and the retained welding points are re-clustered, and the remaining welding points are discarded.

8. The bonding machine welding quality detection method based on machine learning as claimed in claim 5, characterized in that: In step S300, if false alarms occur frequently or the actual welding quality is poor but no alarm is given, the input quantity of the cluster is adjusted based on the distribution of cluster points in the cluster coordinate system; wherein the input quantity of the cluster is the field radius and the threshold.

Images