Copyright © 2021-2022 SHENZHEN KINGFORD TECHNOLOGY CO.,LTD. All Rights Reserved
At first glance, artificial intelligence seems to be a fairly simple concept - it is a device for performing tasks that are usually performed by humans.
Machines have been helping people for a long time, and most handicrafts are now done on machines. The most common replacement has occurred on production lines, where machines perform tasks originally assigned to humans, but with faster speeds, 24/7 operation and lower costs. These machines are programmed to perform specific tasks and do not have their own "intelligence" in the traditional sense, they just get the job done very quickly and reliably.
"Is the apple ripe?" is a trivia question for a human, but in order to get an answer from a machine, it must first be trained.
Thus, artificial intelligence means a system capable of making decisions, but, as in the case of the people it replaces, it needs to be trained.
How to train a machine and how to make this knowledge useful for solving specific problems? The term "artificial intelligence" is immediately reminiscent of robots from sci-fi films such as I Robot, and just like in this film, the artificial intelligence system contains many fragments. One of the best analogies to illustrate the work of artificial intelligence is the relationship between an athlete and a coach.
In the modern world, artificial intelligence has many applications. Its main purpose is to repeat and optimize the processes that a person performs using their five senses. It is these five senses that AI systems recreate, although more often than not, only one is needed to solve a particular problem.
Since vision is one of the main senses that determines human behavior and interactions, it is not surprising that machine vision applications have become the main focus of AI development.
Other applications include processing information from sensors - simulating touch, for example, to detect problems caused by excessive vibration of equipment, simulating hearing for voice analysis, and evaluating sounds that sound "wrong". Taste and smell will soon appear on this list, so the creation of the first perfume by a computer is only a matter of time.
Automatic optical inspection is used to check the quality of the soldering and the correct position of the parts. At the same time, automated PCB inspection is much faster and more accurate than manual inspection.
Given the small size, the large number of small parts, and over a thousand solder points on PCBs, manual control becomes impossible. In addition, the demand for printed circuit boards is growing rapidly, so in order to be competitive and meet tight deadlines, it is necessary to reduce production time.
Automatic optical control monitors the quality of soldering after it is completed on the production line, so errors can be detected early. Correcting any error later in the process will result in higher costs.
The automatic optical control system checks the location of the components, their size, polarity; barcode and the presence of missing elements, soldering points are also checked. One difficulty is that component sizes and other specifications can vary from board to board.
Accurate PCB quality assessment requires the use of special cameras and a powerful AI system that uses advanced algorithms to process visual information quickly and accurately, allowing even faster data processing and more reliable results.
The main objectives of all industries are to increase output, reduce costs, reduce time to market and minimize manufacturing process deficiencies. There is always a need to work better, faster, more productively.
The main task of AI is to enable machines to make decisions or make choices based on input data, mostly visual. These inputs often come from specialized hardware that lacks the processing power to make smart decisions or consider various factors.
The implementation of these capabilities in existing systems requires high costs for upgrading and installing new hardware. Bringing these processes to the cloud simplifies this task, so with the advent of new technologies, there are many new options.
The main limitation remains the computing power required to implement artificial intelligence technologies. Such resource-intensive image processing algorithms put a significant load on the computer's central processor and sometimes cannot be implemented at all using existing technologies.
Previously, the upgrade process was difficult because GPU boards were not designed for such tasks. Although they provide great performance and help to increase the processing power of the system, this is not the most efficient approach to solving the problem.
Thanks to the machine learning revolution and high definition image processing, quality control is becoming fully automated in many industries. It is no longer necessary to perform a manual visual inspection of selected units, as new image processing algorithms are able to detect manufacturing defects, checking each individual unit. And the same process is now also used to detect counterfeit products in the supply chain.
Machine learning (ML) algorithms use huge amounts of data extracted from a specific domain to optimize processes and achieve more efficient results. ML systems work by training to recognize patterns and correlations that connect data points. This requires robust algorithms, large datasets of relevant information, a narrow scope, and a concrete goal.
According to some of the best companies and most established researchers in the field of machine learning, deep learning is an innovative technology that can be applied to revolutionize hundreds of industries. Much of the abstract work has been done and the time has come to use machine learning and transform it into sustainable business models.
For image recognition, it is critical to correctly identify all images in order to distinguish those relevant to the dataset. The more examples of the corrected images an ML-based system is exposed to, the more accurately the system will detect patterns and identify differences in the field.
Although automated visual inspection of PCBs has made significant progress in recent years, several studies show that deep learning surpasses traditional machine-based classification and feature extraction algorithms.
In most PCB manufacturing facilities, defects are initially detected by an automatic inspection machine (AOI). Secondly, a quality control engineer checks each reported PCB.
Unfortunately, the AOI machine can mistakenly classify many cards as faulty due to a scratch, a small hole, or the presence of nanoparticles such as dust, paper scraps, or tiny air bubbles. In addition, even skilled engineers can make mistakes during inspection operations. Consequently, it is imperative to use properly trained ML systems to improve the accuracy of quality control procedures.
Last year, researchers from Yuan Ze University in Taiwan published a study in which they claimed to have achieved a 98.79% PCB defect detection accuracy using YOLO (You-Only-Look) convolutional neural networks.
Compared to other simple classifiers, the YOLO system is widely used in practice. It is a simple and unified object detection model, which can be trained directly using full images. Fast YOLO is the fastest general purpose object detector. YOLOv2 offers the best compromise between real-time speed and accuracy for object detection compared to other detection systems for various detection data sets.
According to the article, the team used 11,000 images,a network of 24 convolutional levels and two fully connected levels. To train the Tiny YOLOv2 algorithm, the researchers used the Keras framework running on an Nvidia TITAN V GPU.
In 2019, the Organization for Economic Co-operation and Development (OECD) reported that counterfeit goods and imitations accounted for 3.3% of global trade volumes. This percentage is constantly increasing every year. The country most affected by the counterfeiting business is the United States, whose trademarks or patents have been linked to 24% of the counterfeit products seized.
Electronics is one of the sectors most affected by counterfeiting. As the price of critical components rises and supply chain disruptions occur, many fake products enter the market to fill the gaps left by the original products. To make matters more difficult, most counterfeit components are nearly indistinguishable from the originals.
Counterfeiting of finished products, especially consumer products, can be detected through a visual inspection. However, it takes a trained eye and considerable knowledge of the original products to distinguish between a fake and an authentic item.
ML algorithms can analyze millions of images and detect even the smallest inconsistencies and anomalies in shape, color, texture and size.
When it comes to electronic components, it becomes more difficult to use visual inspection alone. The market is full of counterfeit units and inventory. Using counterfeit or second-hand components could cause damage to equipment and property, product recalls, and potential liabilities.
In addition to image recognition, other ML methods and algorithms are needed to improve the accuracy of detection of counterfeit or modified parts. In recent years, blockchain technology has quickly become one of the key solutions for protecting the authenticity of products and components and securing the supply chain.
For more information about PCB used in AI, please feel free to contact us.