Kang generation technology: AI and its application in PCB AO1 optical detection

 Kang generation technology: AI and its application in PCB AO1 optical detection

  Author introduction: xing-jun gao, engaged in PCB industry for 20 years, in the field of AO1 automatic optical inspection from 18 years, the current health commissioner, generation of smart technology products, familiar with the process of PCB, the principle, application and development trend of AO1 understanding and research, has published, "big data applications in AO test", "3 d printing application in PCB manufacturing and The application of DFT in AOI

    Permeability shipKang generation technology: A1 and itsThe application of optical detection in PCB AO1The introduction20 years ago to join the PCB industry, familiar with PCB orientation is the direct mount guard of each process, the special deep impression was worked as an intern in electroplating line. Because of the poor working conditions, personnel loss fast, so the plating line "practice" the longest time. Every cylinder before plating line hanging timer, will need to be set according to the artificial time hanging basket of PCB from mentioned under a cylinder, a cylinder until the entire electroplating process. Most of the other processes of up-down material also is artificial, the etched lines after artificial visual quality inspection. Now, with the continuous increase of labor cost, PCB manufacturing, most of the physical labor or even a small number of mental work has been replaced by mechanization, electrification, automation and information technology. And in the implementation of intelligent, found that most of the PCB manufacturing equipment cannot fully support intelligent requirements, the corresponding transformation mainly for automation. There is currently no unified interface specification PCB manufacturing equipment, semiconductor equipment parts manufacturers reference interface specification, but because of its high cost, temporarily unable to widely used in PCB manufacturing. The PCB manufacturer designed according to their own needs in their respective specification, the status quo is multifarious.The white parts

  AI though beyond human in certain areas, impress people a few times of man-machine war: 1997"Deep blue" beat world number one international chess masters; In 2006, five Chinese chess grandmaster eventually on the super computer "wave tissot" hands. In 2011, "Watson" in the quiz show jeopardy against two human champions; Especially in 2016's "alpha go" victory over the champions league in the world But these all belong to the special customized task of artificial intelligence, the learning process is by trial and error to end all possible move, Al showed remarkable feats in the computing power and storage capacity. Human intelligence is different, its the memory - forecasting model, different area of the cerebral cortex (visual, auditory and somatosensory) have the same, a powerful general algorithm, the ability to predict the future is the key to human intelligence. When PCB practitioners attempt to AI technology to the ground, not because the vogue of A1, but need to use the AI technology to solve the practical problems in PCB manufacturing, further improve the production efficiency and product quality or even replace part of knowledge workers. What A can do in AOI process? Can be applied to which links such as the problem is to discuss the content of the next?

1. 

   AIArtificial intelligence is the study of make some of computers to simulate human thinking process and intelligent behavior, such as learning, reasoning, thinking, planning, etc.) of the subject, is a branch of computer science, the research mainly includes the robot, in the field of speech recognition, image recognition, natural language processing, simulation system and expert system, etc. When people talk about artificial intelligence, Machine Learning (ML, Machine Learning), Deep Learning (DL, Deep Learning), depth of neural networks (within DNN, DeepNeural Network), the convolutional neural Network (CNN, The concept of the Convolutional Neural Network) is often mentioned. AI is the goal of the people, machine learning is the main route, to realize the A1 machine learning algorithm with linear regression, logistic regression, integration methods, support vector machine (SVM), neural network, and deep learning, etc., and machine learning can be divided into the supervision according to whether there is a label, a semi-supervised learning, unsupervised learning and reinforcement learning, structured learning and migration.Deep learning is one of the most commonly used in machine learning algorithms, especially for the application of computer vision, the depth of the neural network is to imitate the brain mechanism of a deep learning method. The following will discuss depth, convolution, neural network respectively.

2. 

   The convolution operationConvolution operation: the definition of convolution algorithm is widely used in signal and linear system, digital signal processing and image processing system, from the definition of mathematical formula (1), convolution is a function (unit response or convolution kernels) in another function (the input signal) on a weighted base.

   In the field of computer vision, convolution kernels defined patterns, convolution operation is calculated in each position and the pattern of similarity degree, the current position and the model is more like, the stronger the response. Convolution kernels is usually smaller size of the odd number matrix, digital image is a relatively large size of the 2 d (multidimensional or multi-channel feature maps) matrix, as a convolution operation, in the form of sliding window, from left to right, from top to bottom, each channel is multiplied by the corresponding position of the sum. If the convolution kernels as Weight (Weight), and into a vector for w, image corresponding to the location of the pixels into a vector for x, then the position of the convolution results can be represented by formula (2), the vector inner product + offset (Bias)

   From the Angle of the function to understand the convolution operation above, the following through the two familiar application example to further illustrate. Figure 1 is an early morphological detection logic that is used by the AOI prefab detection template matrix (convolution kernels) sliding on scan images, when the graphics meet template defined features, will generate the corresponding characteristic information, and then compared with standard graphic feature information to find the fault location. According to different defect point form, the AOI need to define different detection logic (matrix) template, such as T, Y, L, K and H, AOI engineer requires constant iterative detection template matrix parameters optimized, good design of convolution kernels, will produce good results. Figure 2 is commonly used in image processing of gaussian image smooth, you can see on the right side of the processed image grey value distribution more uniform, filtering the noise in the image is smooth. Convolution algorithm using the key lies in the design of the convolution kernels, the main role in image processing: image preprocessing and feature extraction, the characteristics of image output to the next link analysis and understanding.

3

   the depth of the neural networkNeural network is a kind of application is similar to the structure of the brain synapses connection information processing, the mathematical model is also a kind of inspired by biological programming paradigm, can let the computer to study of observation data, find out the optimization approach to solve the problem.Artificial neural network to learn the concept of biological neural networks are two very important - cell and the connection weights. Next, in the familiar examples of "AOI equipment evaluation" to understand the workings of a neuron.Hypothesis: a PCB manufacturers need to purchase AOI equipment, AOI engineer usually according to the following factors to decide whether to buy a supplier of AOI equipment: detection and false point rate, material and process adaptability, operation simplicity and capacity, etc. The assessment program did not arranged according to the weight, different PCB manufacturers focus is different, the weight have different Settings, such as A high-end product manufacturers, with special material and complex process, it will set higher weight value for detecting ability and adaptability; And B manufacturers mid-range products, it may focus on operational simplicity and scanning speed. What projects (or called input, predictive factors, characteristics, these are the concept of machine learning) need to assess, according to their respective actual circumstances, this process is features extraction or engineering. Will evaluate the project and the corresponding weight value as input into the neural model of figure 3, can be calculated according to the linear function of the corresponding eigenvalues, then a nonlinear activation function is 0 to 1 (or 1 to 1) between a certain number, thereby realizing logic operation of biological neurons. As the change of weight and bias, can get different decision-making model. In depth study, the weight w and bias b derived from data driven, and the weights and bias is the convolution kernel parameters.tRt 

   Neurons are activated)Figure 3: the neuron modelIf each neuron output of the simulation of biological neural network connection to the next neuron input, can form of artificial neural network (see figure 4), in case the evaluation project "false point rate" in the input, determined by the output of neurons at the next higher level, in order to predict the results, need according to the input neurons at the next higher level and the connection weights to make decisions, such as: 0 if adopting multiple partitions Detection engine to filter the non-critical region of non-critical defects; 2 whether to adopt full spectrum light source to ensure access to clear images; (3) whether to adopt non-contact linear motor to ensure smooth moving object to be tested so that to obtain the stable image, etc.Deep learning, is a powerful collection of numerous learning algorithm for neural network learning. In a broad sense is the process of solving the relationship between the input and output, in a narrow sense is the process of solving the neuron weights and bias. Neural network according to the input layer, hidden layer and output layer depending on the type of network to connect the depth of the neural network, the hidden layer of the layer number and function of module decides the depth and type of neural network, such as the FNN, CNN, RNN and GAN, etc. Deep learning is the process of the training set first labeled data input to the neural network, after each layer neural network processing, so as to minimize the error between the output and the expected value and the loss function minimization (loss function is used to measure the actual behavior and expected behavior deviation), the process is mainly through the forward propagation, the BP algorithm, the loss function to iterative update power The weight and bias. In neural network training process often encounter gradient explosion, training slowly disappear or gradient and the fitting problem. The topic is too big, not discussed here.

3.

   The application of AIArtificial intelligence system mainly consists of three parts: 0 information input. Through a variety of sensing equipment to detect the dynamic changes of the physical world, thus acquiring a large amount of data; (2) the decision-making process. Will get a large amount of data is applied to machine learning model of reasoning, prediction, or decision; (3) performs the output. According to the result of reasoning or prediction to perform the corresponding action. In short is to a large number of input data through the regression, such as integrated machine learning algorithms to establish the prediction model, will build a good model is applied to the actual data sets to forecasting results are obtained. An AIE is widely used in financial, medical, education, public security, transportation, communications, agriculture, meteorology, and services in areas such as, table 1 lists some common application scenario A.

   AO1 optical detectionThe brief introduction of the basic concepts of AI, and mentioned the AI is commonly used in computer vision algorithm, these visual algorithm has been widely used in the AO1. AOI evolved by the artificial visual automatic optical inspection, the working principle is: first of all, through visual algorithm on the CAM data of standard "learning" to the desired image feature information, and then on each piece of PCB scan images used to learn good model in the training set for feature extraction, compared with standard data have the characteristics of the images, according to the given rule (detection Standard) report points out the need to detect problem. AO1 since, as a typical application of computer vision, there is the same as the computer vision.

    The visual obstacle to AOTImaging process information loss: when people try to understand an image, previous experience and knowledge will be used for the observation of people understand the process of image in the unconscious usually is completed. The need to involve math computer vision, pattern recognition, artificial intelligence, psychological physiology, computer science, electronics, and other subjects in the field of the results and methods. So, for AOI, due to the PCB board of 3 d scene onto a 2 d space lost a lot of information, especially the depth information, such as illumination, material properties, orientation and distance information is reflected into the only measured values - grey value. The same 2 d plane projection may be produced by the infinite possibility of 3 d scene projection, thus the inverse process from 2 d to 3 d is a pathological process, qualitative problems or discomfort, observation data are insufficient to constraint problem of the solution, so to take advantage of a priori knowledge or the introduction of appropriate constraints.Such as in the AO [tests, often encounter scan images for open circuit (2 d) image, but in fact to be true, open or oxidation point on the line, residue, dust (3 d)...Image block: in computer vision, image block is usually refers to the light shade, physical barrier, since the shade or mixed shelter, because the image block not only lost part of the target information, and the introduction of additional interference. Shade in the AOI is real the disadvantages of the foreign body cover (see figure 5), the short circuit are covered with sticky foreign body, using tools can be seen in a copper after cleaning, but cannot completely remove foreign body, need to use fiber brush wipe. So, AOI through gray judgment under the foreign body whether there is defect.

     Local window with a global view: usually image analysis algorithm to analyze the operation is a specific storage unit in the memory (a pixel in the image) and its adjacent cell, only from the local or only some local holes can obtain image, image is usually very difficult. AO1 specified width according to the different resolution scan, and divided into specified the size of the image block, so AO1 detection algorithm and local analysis processing, not like E - Test to join the PCB network analysis, only will join in the logic to handle FuZhuCeng for functional analysis.

4.

  AOl Ф AI applicationsAs the key part of quality control in PCB manufacturing process, AOI process, maintenance on the validation process needs more artificial participation, operators need to press the false faults, repair, scrap faults are classified, and corresponding repair or mark on board action, at the same time record quality reports. In these link, how to improve the production efficiency, reduce production costs (especially Labour costs), and reducing quality anomalies caused by human factors (chafed in the process of handling, misjudgment and false negative in the confirmation process, etc.) is a concern of the industry.As mentioned earlier, AlI technology application is the hope can use the simple and practical method to solve practical problems, unlike today's intelligent lighting, traditional lighting can simply press the switch can turn on/turn off the lights, intelligent lighting needs out his phone and open the App, open the lamp and adjust the color brightness or long through voice control, the process has become more complicated. In the process of integrating A] technology to AO detection, the need to avoid similar to Al and AI behavior, but to form A complete set of valuable solutions. Below is divided into upper and lower material, AOI inspection and CVR confirm three parts to discuss.

5.

   AI applicationsAs the key part of quality control in PCB manufacturing process, AOI process, maintenance on the validation process needs more artificial participation, operators need to press the false faults, repair, scrap faults are classified, and corresponding repair or mark on board action, at the same time record quality reports. In these link, how to improve the production efficiency, reduce production costs (especially Labour costs), and reducing quality anomalies caused by human factors (chafed in the process of handling, misjudgment and false negative in the confirmation process, etc.) is a concern of the industry.As mentioned earlier, AlI technology application is the hope can use the simple and practical method to solve practical problems, unlike today's intelligent lighting, traditional lighting can simply press the switch can turn on/turn off the lights, intelligent lighting needs out his phone and open the App, open the lamp and adjust the color brightness or long through voice control, the process has become more complicated. In the process of integrating A] technology to AO detection, the need to avoid similar to Al and AI behavior, but to form A complete set of valuable solutions. Below is divided into upper and lower material, AOI inspection and CVR confirm three parts to discuss.

6.

   AO1 detectionMainly includes the AOI equipment operation and the logic operation, AOI application system has been simplified the operations (material obtain, registration, light, calibration and application set), a key to achieve the basic operation. Parts manufacturers adopt the way of scan qr code to need to scan the material number to be obtained to ensure fast accurate. So for now, AO1 operation process temporarily without AI technology optimization. If forced to adopt the visual system or speech recognition system to AOI system operation, the process will become more complicated. AO1 detection logic after years of iteration and optimization, according to the training process of deep learning evaluation, is the optimum detection model. As for logic operation efficiency, and AI, AO vision algorithms are used quite a lot of matrix multiplication and convolution operation, because gpus can efficiently handle matrix multiplication and convolution operation, predictably, the GPU will increasingly used for AOI, in order to improve the efficiency of the logical operation. So after AO1 processing is a key link in the process of application of AI technology, namely the AI technology is applied to the maintenance station confirmation in order to reduce investment of equipment and manpower cost,

7

  CVR confirmationThis process is to focus on the most part, mainly includes the false points filtering and classification processing of two parts. If it can reduce the false points rate accordingly reduce the board handling, reduce maintenance equipment investment and reduce the labor costs of maintenance, the fake here does not mean that the logical false points (without any exception report faults), but don't want to get the report out shortcomings, such as dust, no sticky foreign body and oxidation. According to statistics, this kind of "fake" accounted for 30% of the total number of faults is less than 80% (the ratio according to the manufacturing equipment of different manufacturers, production process control and environmental factors such as different will have larger difference). Common false spot diagram as shown in figure 6

    conclusionThough the AI technology has been in AOI is widely used in image processing, but the repair link still has a huge application scenarios and space, especially in today's AI fast iterative algorithm, the future maintenance system will be integrated intelligent inspection system of various kinds of AI algorithms.Of course, these A1 algorithm can be simple regression, classification algorithm, can be complicated reinforcement learning, learning, structured strong ai and the weak ai is today's controversial philosophical questions.