CN108615075B - Automatic parameter adjusting method - Google Patents
Automatic parameter adjusting method Download PDFInfo
- Publication number
- CN108615075B CN108615075B CN201810439606.3A CN201810439606A CN108615075B CN 108615075 B CN108615075 B CN 108615075B CN 201810439606 A CN201810439606 A CN 201810439606A CN 108615075 B CN108615075 B CN 108615075B
- Authority
- CN
- China
- Prior art keywords
- parameter
- wire bonding
- impact
- pressure
- delta
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Hardware Design (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Geometry (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Wire Bonding (AREA)
Abstract
The invention relates to an automatic parameter adjusting method which can be applied to ultrasonic wire bonding equipment for semiconductor packaging, and comprises the following steps: generating an initial parameter combination according to a historical database, wherein the parameter combination comprises one or more of impact pre-pressure, impact time, impact pressure, wire bonding time, wire bonding pressure and wire bonding energy; the ultrasonic wire welding equipment executes and measures the quality coefficient of the product according to the initial parameter combination; feeding back and updating the parameter combination according to the product quality coefficient; and repeating the steps until the product quality coefficient is all qualified. The method takes scene information and target effect (quantification) as 'input' and 'result' respectively, and takes target optimization parameters (quantification) as 'built-in parameters' of model optimization. In practical application, a state space model can be generated by collecting a small amount of data on a generation site, and meanwhile, required optimization parameter combinations are searched through a VNN algorithm based on a back propagation algorithm.
Description
Technical Field
The invention relates to an artificial intelligence deep learning technology, in particular to an automatic parameter adjusting method.
Background
An important application scene of the wire bonding machine is in the LED industry, the wire bonding machine is high in technical difficulty at present, only a few manufacturers can produce the wire bonding machine, a main manufacturer abroad is KS, and the automation of laser and green waves of a large family is domestic. Imported equipment still occupies the domestic major market, especially the high-end market. One of the main problems of domestic equipment is the consistency problem, that is, the difference between products produced by the same equipment under the condition of the same parameter selection is large, wherein some equipment can not reach the quality inspection standard, and the parameter needs to be adjusted.
For the quality detection of the wire bonder product, a welding spot, two welding spots and a wire arc are mainly considered. The consideration factors of a solder joint mainly include solder ball thrust, solder ball size, solder ball shape, burr ball and the like. Among them, the thrust of the solder ball is one of the most critical standards, and an excessively small thrust value indicates that the solder ball is not well adhered to the electrode, and the reliability of the product is risky. Typically greater than 30 grams is required but not too large, which can cause other problems, and there is typically a recommended target optimum, such as 45 grams for alloy wire and 35 grams for copper wire. The measurement of the thrust is detected by a special thrust gauge, and the height of the push broach is more than 5 um. The consistency of the thrust of a welding spot of a product is not good under the condition of the same welding parameters, and the problem is urgently needed to be solved by domestic equipment.
Disclosure of Invention
The invention aims to solve the technical problems of quality detection of the products of the wire bonding machine in the prior art.
In order to achieve the above object, the present invention provides an automatic parameter adjusting method, which can be applied to an ultrasonic wire bonding apparatus of a semiconductor package, and comprises the following steps:
generating an initial parameter combination according to a historical database, wherein the parameter combination comprises one or more of impact pre-pressure, impact time, impact pressure, wire bonding time, wire bonding pressure and wire bonding energy; the ultrasonic wire welding equipment executes and measures the quality coefficient of the product according to the initial parameter combination; feeding back and updating the parameter combination according to the product quality coefficient; and repeating the steps until the product quality coefficient is all qualified.
Preferably, the step of updating the parameter combination according to the product quality coefficient feedback includes: inputting an initial parameter X vector based on an MLP state model, and calculating output thrust; finely adjusting each input parameter delta x, and respectively calculating the change delta y of the final thrust brought by the change of each parameter; calculating the gradient change delta G as delta y/delta x on each parameter dimension; calculating a distance value (error) epsilon from the output thrust to the target thrust; calculating the parameter updating amount of each input dimension based on the gradient change Delta G under the current distance epsilon: d ═ epsilon/. DELTA.g; updating the initial parameter Xnew ═ X + α D by the usual learning rate α (0.01); and repeating the 6 steps until epsilon is smaller than a threshold value theta which is 1, and outputting the latest X vector, namely the optimized parameter combination obtained by the VNN searching method.
The invention extracts a back propagation algorithm from a deep learning standard process, invents a Virtual parameter Network (VNN) without a specified structure, takes scene information and target effect (quantization) as an input and a result respectively, and takes a target optimization parameter (quantization) as a built-in parameter of model optimization. In practical application, a state space model can be generated by collecting a small amount of data on a generation site, and meanwhile, required optimization parameter combinations are searched through a VNN algorithm based on a back propagation algorithm. In addition, the invention helps the curve overtaking of the domestic welding wire machine equipment manufacturer to overtake the welding wire machine equipment manufacturer which is temporarily in the leading position internationally by combining the deep learning artificial intelligence with the automatic production equipment.
Drawings
Fig. 1 is a schematic flow chart of an automated parameter adjusting method according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of an embodiment of the present invention.
Detailed Description
Other features, characteristics and advantages of the present invention will become more apparent from the following detailed description of embodiments of the present invention, which is to be read in connection with the accompanying drawings.
The full-automatic ultrasonic wire bonding machine (hereinafter referred to as wire bonding machine) is generally applied to the inner lead welding of high-power devices such as Light Emitting Diodes (LEDs), laser tubes (lasers), medium and small power triodes, integrated circuits and some special semiconductor devices. The wire bonding machine is a complex optical, mechanical and electrical integration device, and generally integrates computer control, motion control, image processing and network communication. The wire bonding machine realizes surface welding of different media through physical change, ultrasonic waves (generally 40-140 KHz) generated by mechanical vibration energy of ultrasonic frequency are utilized, a lead and the surface of a metal to be welded rub with each other, a surface oxide layer is broken, and firm metal bonding is formed on a welding surface. Effectively avoids the phenomena of splashing, oxidation and the like generated during welding.
Deep learning (deep learning) is an algorithm that performs high-level abstraction of data with multiple processing levels, using processing projects that contain complex structures or consist of multiple nonlinear transformations. An artificial neural network is a common deep learning model, which is a mathematical or computational model that mimics the structure and function of a biological neural network, and is used to estimate or approximate functions. Deep learning neural networks, such as convolutional neural networks and deep confidence networks and recurrent neural networks, have been widely used in the fields of computer vision, speech recognition, natural language processing, audio recognition, and bioinformatics. The deep learning model, whether it is structurally complex or not (from a three-layer fully connected network to hundreds of layers of ResNet), is based on a back propagation algorithm (backpropagation) developed and practiced by 1986Rumelhart, Hinton, Williams et al for deep learning neural network learning training.
Back-propagation algorithms, also known as "error back-propagation," are commonly used in conjunction with optimization methods (such as gradient descent) to train artificial neural networks. Generally, a supervised learning method is a generalization of the Delta rule of neural networks, and the gradient can be calculated for each layer iteration by using a chain rule. Back propagation requires that the excitation function of the artificial neuron be minute. The method calculates the gradient of the loss function for all weights in the network. This gradient is fed back to the optimization method for updating the weights to minimize the loss function. Back propagation requires a known output to be obtained for each input value to calculate the gradient of the loss function.
However, the known deep learning model is used in the field of industrial parameter adjustment (wire bonding machine) to find the optimal parameter combination, and faces the following three problems:
1. the deep learning model is mainly based on historical data learning aiming at relatively fixed scenes/environments: the industrial equipment has a complex structure, the influence factors are various in the production process, the result of equipment of the same model is different under the same parameters, and the shutdown and the restart are also changed, so that the model trained based on historical data is poor in performance, each scene is changed in practical application, and the historical data of the completely same scene is not available for reference.
2. Traditional model training requires a large amount of data (multidimensional, with known results): data collection in the industrial field is still imperfect, data collection cost is high, and in practical scene application, data collection in new equipment or a new scene of a wire bonding machine needs to be completed in limited rounds (only less than 10 data points at minimum).
3. The parameters are inputs, the optimization is a process, the goal is accuracy: the traditional deep learning function is mainly recognition, classification and prediction, the model is a deep learning algorithm with definite mathematical expression, the accuracy of the model is maximized by optimizing built-in parameters, and the function applied by the parameter adjusting type is the optimization of the parameters, the model is a real device/production flow, the determined mathematical expression is difficult to establish, and the parameters to be optimized are optional parameters of actual parameter adjustment. For industrial tuning, accuracy is an input, optimization is a process, parameters are targets, but there are no a priori parameters as known results in the new scenario.
Fig. 1 is a schematic flow chart of an automated parameter adjusting method according to an embodiment of the present invention. As shown in fig. 1, the method comprises the steps of:
step one, generating an initial parameter combination according to a historical database. Specifically, according to scene information such as wire rods, equipment models, magnetic nozzle specifications, chip types and the like, a group of closest parameter combinations are matched in a database, and the parameter combinations mainly comprise 6 most important parameters influencing a welding thrust: 1. impact pre-pressure; 2. impact time; 3. impact pressure; 4. wire welding time; 5. wire bonding pressure; 6. the bonding wire energy and these parameters correspond to a reasonable variation range (upper and lower limits of the tuning parameters).
The VNN method adopted by the embodiment of the invention is a method for searching new parameters under the condition that the initial parameter combination can not reach the standard. The application of VNN needs to satisfy two preconditions for the state space (mathematical mapping of state space parameters to quality control targets (thrust magnitude)): 1, the state space is continuous, and in the continuous space state, a mathematical model can be constructed according to correlation analysis such as reciprocal slope of measured data, and the direction of the next search is deduced, which is a precondition that the VNN can be applied. If the state of each time has no correlation with the position searched in the previous time in the discontinuous space state, the position of the state meeting the qualification of the quality inspection may appear in any area of the space, and only random search can be performed. 2 the state spaces are similar. Under the similar state space, historical data and experience have reference values, and the method can be used for generating initial parameter combinations and enhancing search algorithms. If the state space is large in difference, each parameter adjustment needs to be started from zero, and the limit of the number of times of iteration is increased.
And step two, the ultrasonic wire welding equipment executes and measures the quality coefficient of the product according to the initial parameter combination.
And step three, feeding back and updating the parameter combination according to the product quality coefficient.
Specifically, the VNN algorithm of the embodiment of the present invention includes two components: the method comprises the following steps that 1, an MLP-based multilayer perception neural network is used for building a state space mapping model to search a mathematical mapping relation between an adjusting parameter and a target thrust; and 2, a parameter search algorithm (VNN search) based on a back propagation algorithm is used for searching for proper parameters, and an output result obtained after the proper parameters are input into the MLP state space model is the target thrust.
The MLP state space mapping model method is as follows: a small amount of data (a minimum of 3 data points) is collected in the field, each data point comprises a set of parameters and a thrust measurement result, and an MLP multilayer perceptive neural network is adopted. The basic structure of the MLP network is as follows:
the multilayer perception neural network model in deep learning is more suitable for the scene of the variable data, so the architecture flow of the model of the invention is as follows:
the multilayer perception neural network comprises 10 neurons in a hidden layer except for an input (6 units) and an output (1 unit, corresponding to a thrust value). The model liberalization parameter matrix is 88 weights/bias.
The realization platform is Python, adopts the Anacoda3(64bit) environment, has called the keras (tensoflow back) module, numpy module.
The hidden layer activation function is tanh and the output layer activation function is linear.
Performing online learning according to newly collected data each time, wherein the training method is a standard Adam post-propagation training method, the cost function is a prediction variance value, the training period is 500 cycles, and the learning rate is 0.01
Model training objective function prediction accuracy
Wherein Y is the actually measured thrust force,
to predict thrust.
It should be noted that the VNN search method is as follows: based on the MLP state model, any set of parameters can be projected to a unique thrust value through a mathematical mapping, but can only be solved through a mathematical relationship because the parameter dimensions (such as 6 parameters related to thrust) exceed 3 dimensions. After the mathematical mapping relation is determined, the method for searching qualified parameter combinations in the multidimensional space is a standard reverse conducting learning method:
inputting an initial parameter X vector based on an MLP state model, and calculating output thrust;
finely adjusting each input parameter delta x, and respectively calculating the change delta y of the final thrust brought by the change of each parameter;
calculating the gradient change delta G as delta y/delta x on each parameter dimension;
calculating a distance value (error) epsilon from the output thrust to the target thrust;
calculating the parameter updating amount of each input dimension based on the gradient change Delta G under the current distance epsilon: d ═ epsilon/. DELTA.g;
updating the initial parameter Xnew ═ X + α D by the usual learning rate α (0.01);
and repeating the 6 steps until epsilon is smaller than a threshold value theta which is 1, and outputting the latest X vector, namely the optimized parameter combination obtained by the VNN searching method.
And if the search result fails to meet the requirement, combining the new measurement result and the old result, and training the MLP state space model again on line and carrying out VNN search.
And step four, repeating the steps until the product quality coefficient is all qualified, and if the welding spot thrust is close to within 10% of the target value, the product quality coefficient is qualified.
The embodiment of the invention extracts a back propagation algorithm from a deep learning standard process, and invents a virtual parameter network without a specified structure, takes scene information and target effect (quantification) as 'input' and 'result' respectively, and takes target optimization parameters (quantification) as 'built-in parameters' of model optimization. In practical application, a state space model can be generated by collecting a small amount of data on a generation site, and meanwhile, required optimization parameter combinations are searched through a VNN algorithm based on a back propagation algorithm.
In one embodiment, the parameter adjustment operation of the wire bonding LEDs is performed on a new automated wire bonding machine in a wire bonding plant, and the actual measurement results show that the algorithm results reach 90.2% accuracy within 7 data points suggested by the 3 round parameters (two adjustments), as shown in the table below and in fig. 2.
| Device | First welding wire machine |
| Product(s) | LED |
| Wire rod | Alloy wire (0.9um) |
| Magnetic nozzle | GAISER1572A-13TR |
| Protective gas | Is free of |
| Chip welding spot | Right side of the |
| Target thrust | 45g |
It should be noted that the above embodiments are only used for illustrating the structure and the working effect of the present invention, and are not used for limiting the protection scope of the present invention. Modifications and adaptations to the above-described embodiments may occur to one skilled in the art without departing from the spirit and scope of the present invention and are intended to be covered by the following claims.
Claims (5)
1. An automatic parameter adjusting method is applied to semiconductor packaging ultrasonic wire bonding equipment and is characterized by comprising the following steps:
generating an initial parameter combination according to a historical database, wherein the parameter combination comprises one or more of impact pre-pressure, impact time, impact pressure, wire bonding time, wire bonding pressure and wire bonding energy;
the ultrasonic wire welding equipment executes and measures the quality coefficient of the product according to the initial parameter combination;
feeding back and updating the parameter combination according to the product quality coefficient;
repeating the steps until all product quality coefficients are qualified;
the step of updating the parameter combination according to the product quality coefficient feedback comprises the following steps:
inputting an initial parameter X vector based on an MLP state model, and calculating output thrust;
finely adjusting each input parameter delta x, and respectively calculating the change delta y of the final thrust brought by the change of each parameter;
calculating the gradient change delta G as delta y/delta x on each parameter dimension;
calculating a distance value epsilon from the output thrust to the target thrust;
calculating the parameter updating amount of each input dimension based on the gradient change Delta G under the current distance epsilon: d ═ epsilon/. DELTA.g;
updating an initial parameter Xnew ═ X + α D by a commonly used learning rate α;
and repeating the 6 steps until epsilon is smaller than a threshold value theta which is 1, and outputting the latest X vector, namely the optimized parameter combination obtained by the VNN searching method.
2. The method of claim 1, wherein if the search result fails to meet the requirement, combining the new measurement result and the old result to re-train the MLP state space model online and perform VNN search; the parameters input to the MLP state space model include one or more of impact pre-pressure, impact time, impact pressure, bond time, bond pressure, and bond energy.
3. The method of claim 1, wherein the MLP state space mapping model method is as follows: at least 3 data points are collected on site, each data point comprises a group of parameters and a thrust measurement result, and an MLP multilayer perception neural network is adopted.
4. The method of claim 1, wherein the step of feedback updating the parameter combinations according to the product quality coefficients comprises:
searching for the required combination of optimization parameters by a VNN algorithm based on a back propagation algorithm.
5. The method of claim 1, comprising applying to a semiconductor package.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810439606.3A CN108615075B (en) | 2018-05-09 | 2018-05-09 | Automatic parameter adjusting method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810439606.3A CN108615075B (en) | 2018-05-09 | 2018-05-09 | Automatic parameter adjusting method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108615075A CN108615075A (en) | 2018-10-02 |
| CN108615075B true CN108615075B (en) | 2021-09-28 |
Family
ID=63662546
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810439606.3A Active CN108615075B (en) | 2018-05-09 | 2018-05-09 | Automatic parameter adjusting method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108615075B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110909507A (en) * | 2018-09-12 | 2020-03-24 | 台达电子工业股份有限公司 | Method for suggesting soldering tin process parameters |
| CN109472358B (en) * | 2018-10-17 | 2021-10-19 | 深圳市微埃智能科技有限公司 | Neural network-based welding process parameter recommendation method and device and robot |
| CN109446639B (en) * | 2018-10-25 | 2023-05-12 | 重庆大学 | Detonating cord crimping parameter autonomous optimization method based on convolutional neural network |
| US11501116B1 (en) * | 2021-07-19 | 2022-11-15 | Sas Institute Inc. | Quality prediction using process data |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6556876B1 (en) * | 2000-10-12 | 2003-04-29 | National Semiconductor Corporation | Hybrid fuzzy closed-loop sub-micron critical dimension control in wafer manufacturing |
| CN101807265A (en) * | 2008-09-30 | 2010-08-18 | 洛克威尔自动控制技术股份有限公司 | System and method for dynamic multi-objective optimization of machine selection, integration and utilization |
| CN103959769A (en) * | 2012-02-02 | 2014-07-30 | 松下电器产业株式会社 | Methods and apparatuses for 3D media data generation, encoding, decoding and display using disparity information |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9147166B1 (en) * | 2011-08-10 | 2015-09-29 | Konlanbi | Generating dynamically controllable composite data structures from a plurality of data segments |
-
2018
- 2018-05-09 CN CN201810439606.3A patent/CN108615075B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6556876B1 (en) * | 2000-10-12 | 2003-04-29 | National Semiconductor Corporation | Hybrid fuzzy closed-loop sub-micron critical dimension control in wafer manufacturing |
| CN101807265A (en) * | 2008-09-30 | 2010-08-18 | 洛克威尔自动控制技术股份有限公司 | System and method for dynamic multi-objective optimization of machine selection, integration and utilization |
| CN103959769A (en) * | 2012-02-02 | 2014-07-30 | 松下电器产业株式会社 | Methods and apparatuses for 3D media data generation, encoding, decoding and display using disparity information |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108615075A (en) | 2018-10-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108615075B (en) | Automatic parameter adjusting method | |
| US10518357B2 (en) | Machine learning device and robot system to learn processing order of laser processing robot and machine learning method thereof | |
| CN109472358B (en) | Neural network-based welding process parameter recommendation method and device and robot | |
| US20180260712A1 (en) | Laser processing apparatus and machine learning device | |
| US20070100487A1 (en) | Method and system for virtual metrology in semiconductor manufacturing | |
| CN109035242B (en) | HF-ERW welding state detection method and device based on deep learning | |
| JP2008117381A (en) | Time sequential data analyzer, time sequential data analysis system, time sequential data analysis method, program and recording medium | |
| JP2008305373A (en) | Dual-phase virtual metrology method | |
| CN117047286B (en) | Method for processing workpiece surface by laser, processing system, processor and storage medium | |
| JP2018107315A (en) | Machine learning device for printed board assembling work, control device, industrial machine, assembly system, and machine learning method | |
| CN115863193A (en) | Gold-plated palladium-copper-based bonding wire and preparation method thereof | |
| Azadi Moghaddam et al. | Modeling and optimization of the electrical discharge machining process based on a combined artificial neural network and particle swarm optimization algorithm | |
| Weng et al. | A one-dimensional vision transformer with multiscale convolution fusion for bearing fault diagnosis | |
| Pourmostaghimi et al. | Intelligent model-based optimization of cutting parameters for high quality turning of hardened AISI D2 | |
| CN114897102A (en) | Industrial robot fault diagnosis method, system, equipment and storage medium | |
| CN112699903B (en) | Full-flow visual tracking control method and system for motorcycle engine assembly | |
| Huang et al. | Process optimization of SnCuNi soldering material using artificial parametric design | |
| CN118438039B (en) | Welding method for high-voltage quick-plug connector | |
| Hamedi et al. | Optimizing spot welding parameters in a sheet metal assembly by neural networks and genetic algorithm | |
| CN112462600B (en) | High-energy laser control method and system, electronic equipment and storage medium | |
| CN116038163A (en) | Laser intelligent welding system and method | |
| Lahoti et al. | Recurrent neural networks to model input-output relationships of metal inert gas (MIG) welding process | |
| CN113435670B (en) | Prediction method for deviation quantification of additive manufacturing cladding layer | |
| Liu et al. | Using deep reinforcement learning to guide PCBS welding robot to solve multi-objective optimization tasks | |
| TWI692695B (en) | Feedback system for improving process quality of product and method thereof |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |