From the production line to the starry sky, these AI engineers bring algorithms to the sky

In many technology-related live videos, you may have seen such a category: someone will disassemble a mobile phone into parts, and then frame them as souvenirs. These videos allow us to clearly see the precise structure inside the phone and various small parts such as connectors and camera brackets that are usually not easy to see.

The precision and quality requirements of these parts are very high, and in order to meet these requirements, our manufacturing process has been iterating rapidly.

SIM card connectors and camera brackets are examples. These small parts are sintered by metal powder mixed with binder. Therefore, compared with traditional processes, the parts produced by the traditional process have higher density, precision and surface finish, and have more restrictions on the structural complexity of the parts. Small.

However, when it comes to the inspection process before leaving the factory, these advantages increase the difficulty for quality inspectors to find defects in parts. For example, due to the high reflective properties of parts, product defects will be confused with normal reflection, and the complex structure of parts will lead to some defects from a single Angles are not easy to spot. These properties make their appearance detection extremely difficult. Today, when most of the production links have reached a high degree of automation, the appearance inspection of parts and components has become a bottleneck for the production capacity of many manufacturing enterprises.

If you are the person in charge of a 3C parts production company, the orders received this year have increased by 20% compared with last year, which is theoretically a very happy thing.

But in fact, behind the growth of orders, there is another side.

This extra 20% may become a "burden" if the quality inspection process - especially the appearance inspection - cannot be automated. Because you need to recruit more people to complete the extra quality inspection work, and the reality is that it is difficult and expensive to recruit.

Appearance inspection is different from size inspection, which is highly standardized and has a measurement standard that everyone recognizes, so traditional machine learning solutions may be able to solve it. In contrast, visual inspection is much more troublesome: what kind of defects can be tolerated and what kinds of defects cannot be tolerated? , quality inspection workers with a very low error rate) to adjudicate; moreover, due to the irregular shape of the parts, some defects are very hidden, and the quality inspector needs to hold it in his hand to observe from multiple angles, or even press it with his hand to find it. This makes it difficult for general testing equipment and testing methods to meet the requirements.

Can the equipment that can screen out most of the defects be used? In demanding manufacturing, the answer is no.

You must know that once the parts flowing downstream are found to have a small defect, downstream customers (such as assembly plants) will complain, and even return a batch of tens of thousands of products, requiring upstream re-inspection. At this point, upstream parts manufacturers have to pay double or even more labor costs to solve the problems caused by the machine. Therefore, if the machine's capabilities are not up to the mark, manufacturers may not be able to use it more than the gain. Because of this, they often require suppliers of visual quality inspection automation solutions to achieve "zero missed inspections" in the detection capability of certain serious defects (such as cracks).

In the past few years, this "zero missed inspection" requirement has caused many technical teams to fall into the sand. Most 3C parts manufacturers still rely on manpower to complete the highly repetitive work of appearance quality inspection, and bear the loss of missed inspection caused by fatigue of quality inspection workers.

Today's increasingly "difficulty in recruitment" exposes a problem: no one wants to repeat the boring life of "take a sample and look at it, turn it around for 1 minute, put it back, and take another sample and look at it again" day after day. Therefore, it is a matter of time to use machines to liberate the manpower of appearance inspection, even if it is difficult.

Fuchi Hi-Tech, which is engaged in the manufacture of metal powder injection molding (MIM) products, invested tens of millions in related products in 2015. In 2019, it also tried some visual quality inspection prototypes based on deep learning, but the results were not satisfactory. It wasn't until 2020 that the company accidentally got in touch with the Tencent Youtu Lab team that things turned around.

Tencent Youtu is a top artificial intelligence laboratory under Tencent, with profound technical accumulation and industry practice in the field of computer vision. Before sending people to Fuchi Hi-Tech, Youtu has also done some "AI + industrial manufacturing" projects, such as the automatic defect classification (ADC) project co-created with CSOT, which saved CSOT more than 10 million yuan at that time. the cost of.

However, there are significant differences between the two projects. In the appearance quality inspection projects of parts like Fuchi Hi-Tech, they found that in order to prevent the machine from missing parts defects, it is not enough to improve the algorithm's "image recognition" ability, but also need to solve a series of problems such as imaging and subsequent process improvement. , requires developers to know the production line very well.

For example, "dirt" and "bruise" both look like a black spot on the picture, but a small area of ​​dirt may not be considered a defect. How can you ensure that the machine is not wrong? In the process of communicating with the quality inspectors, the Youtu team learned that they couldn't tell the difference if they only looked at the flat image, and they needed to look at it sideways or wipe it to judge. This means that the reason why the machine makes mistakes is that in addition to the algorithm that needs to be polished, the information provided to it by humans also needs to be supplemented.

Therefore, they innovatively applied a "photometric stereo" shooting scheme, that is, adding a circle of light sources to the shooting equipment, allowing the light to illuminate from four angles and taking pictures, and then using these four photos to synthesize a picture with 3D information , showing whether the "black dots" are bumpy or not. In addition, in order to solve the interference caused by inconsistent standards to the algorithm, they also repeatedly confirmed with these photos and "Golden Eye" to fit some "floating standards" in the appearance quality inspection.

Small problems like this crop up in an endless stream of industrial projects, and it can be difficult to come up with a solution without touching the production line in person. As Wu Yongjian, vice president of Tencent Cloud and deputy general manager of Youtu Lab, said, "If we don't step into the game, it will be difficult to do this."

After more than 300 days of on-site research and development, Youtu successfully translated many small problems in appearance inspection into languages ​​that the algorithm can understand, and finally customized the "Tenghui Feitong" AI quality for Fuchi Hi-Tech. detector. The quality inspection instrument can complete the quality inspection of a product in 4 seconds, and can detect 20,000 pieces a day. Serious defects can be zero missed inspection, and only a small amount of data can be added to the appearance inspection of new products. Under the condition that more than 10 equipments continue to be fully loaded, the project is expected to save tens of millions of yuan in costs for Fuchi Hi-Tech every year.

Most importantly, this project helped Tencent Youtu set a benchmark in the consumer electronics industry. After the news spread, more and more manufacturers began to find them, hoping to customize similar products. Therefore, they have subsequently undertaken quality inspection projects of several manufacturing enterprises such as Likai Precision, and have maintained a good result that the missed inspection rate of serious defects is close to 0.

From production line to starry sky

There are too many projects to do, and the pursuit of reusability and standardization will be put on the agenda. Zhang Linliang, an engineer at Tencent Cloud, said that in fact, after completing the Fuchi Hi-Tech project, they have some things that they can use again, such as the photometric stereo solution mentioned above, because the quality inspection of many industries involves dirty, Basic defects such as pits, scratches, etc.

In order to make the accumulation of these industry knowledge more systematic and the reuse more effective, they have classified the contacted products according to the material and process. For example, the middle frame of a certain brand mobile phone and the shell of a notebook are based on "aluminum parts + Anodizing" process, their defect types are naturally similar, and quality inspection products designed according to this idea can be adapted to a wider range of production lines.

Interestingly, these techniques and refined knowledge polished in real scenes can not only be used in the production line of the factory, but also extended to the stars.

At last year's World Artificial Intelligence Conference, Tencent officials announced the "Star Exploration Plan" in cooperation with the National Astronomical Observatory, that is, based on the computer vision technology of Youtu Lab, Tencent's cloud computing and storage capabilities, using "cloud + AI" to help China Tianyan FAST processes the huge amount of data it receives every day and finds pulsar clues through visual AI analysis. Under the same computing power, Tencent helped the National Astronomical Observatory to increase the data processing efficiency by 120 times.

Pulsars are fast-spinning neutron stars produced by stellar evolution and supernova explosions. They can be used for gravitational wave detection, black holes and other related research, which help answer many important physics questions. In addition, because they can emit precise periodic pulse signals, pulsars also have the role of interstellar navigation, so they are called "cosmic lighthouses". It plays a role in interstellar travel similar to the GPS we use on Earth.

After a year of cooperation, so far, the star exploration program has found 22 pulsars from the sky survey observation data. Among them, there are 7 high-speed rotation millisecond pulsars with high observation and research value in astrophysics, and 6 old pulsars with intermittent radiation phenomenon. In addition, the Youtu dynamic spectrum AI model also detected the radio pulse of a magnetostar for the first time.

When summarizing the project experience, Wang Chengjie, the research director of Tencent Youtu Lab, who participated in both the industrial visual quality inspection and the "Star Exploration Program", said that these two things have commonalities in the underlying technology: " Most of the signals received from the universe are the background of the universe, and a small part is related to pulsars. The same is true in industrial quality inspection, most of the parts or the area of ​​most parts are in a normal state , finding defects is also a less probable feature search.”

Therefore, for an institution that has been polishing AI technology for ten years, broadening the application boundaries of the technology and unlocking more scenarios is not as laborious as imagined.

At present, these technologies and knowledge are deposited on a one-stop platform called "Tencent Cloud TI". This platform includes Tencent Cloud TI ONE, TI Matrix, and TI DataTruth three major AI underlying platforms, which can provide algorithms including algorithm development, model A series of development capabilities, such as training, data labeling and data processing, export the research and development achievements of Tencent's excellent laboratories such as Youtu for many years to all walks of life at a low threshold.

At this year's WAIC World Artificial Intelligence Conference, Wu Yunsheng, vice president of Tencent Cloud and general manager of Tencent Youtu Lab, also announced the latest upgrade results of the TI platform - the launch of the AI ​​acceleration function TI-ACC, whose training acceleration performance is higher than that of the native framework. More than 30%, the model inference acceleration ratio is more than 2 times. This will accelerate the landing speed of AI big models in all walks of life.

However, the complexity of real-world scenarios such as factory production lines determines that the standardized path represented by this one-stop platform cannot be achieved overnight. As far as the industry is concerned, Wu Yongjian also believes that it is unrealistic to standardize the entire industrial AI at this stage. What they want to do is to find some subdivision tracks to focus on, such as the standardization of 3C and lithium batteries, and then to unlock new areas.

When looking forward to the future of this platform, Wang Chengjie said that in the next step, it will be highly automated and low-threshold. It will automatically choose which algorithm model to use according to your tasks, and tell you what to do at each step, so as to make more Multiple people can use these abilities.

Epilogue

Talking about manufacturing, "factory relocation" and "South Asian demographic dividend" are all topics that cannot be avoided. Zhang Linliang believes that their work can help deal with such a crisis, because if we achieve artificial intelligence to improve efficiency and allocate manpower to where it is more needed, then our equipment may be cheaper than foreign labor, which helps Consolidate our status as a manufacturing powerhouse.

In recent years, the national level has also encouraged companies to do this. Last month, six departments including the Ministry of Science and Technology announced the "Guiding Opinions on Accelerating Scenario Innovation and Promoting High-quality Economic Development with High-level Application of Artificial Intelligence". The "Guiding Opinions" pointed out that with the promotion of the deep integration of artificial intelligence and the real economy as the main line, with the direction of promoting the opening of scene resources and improving the ability of scene innovation, strengthen the cultivation of subjects, increase application demonstrations, innovate systems and mechanisms, improve scene ecology, and accelerate artificial intelligence. Intelligent technology research, product development and industrial cultivation, explore new models and paths of artificial intelligence development, and promote high-quality economic development with high-level application of artificial intelligence.

Many Internet and technology companies, including Tencent, are working towards this goal, and their researchers are increasingly going out of the laboratory to have more in-depth contact and understanding of the industry. Wu Yunsheng said that he himself has also experienced some changes in his work. In addition to being the general manager of Youtu Laboratory, he is also the head of production and research in the government-enterprise business line. This business line includes industrial energy, operators, cultural tourism , real estate and other industry sectors, which gives him "more opportunities to meet the real needs of these industries."

He said at this year's World Artificial Intelligence Conference Tencent Forum, "Tencent has always firmly believed that technology will better support the high-quality and vigorous development of the real economy with the combination of artificial intelligence technology innovation and industrial scenario innovation. Facing the future, Tencent It will also continue to improve the product innovation capabilities of AI technology and open platforms, and provide enterprises and developers with richer platform-based services.”