Trim: AI + manipulator replaces experienced workers
● Self learning trimming experience by AI
● The robot has the ability of self-learning and self evolution, which can make the trimming speed faster and faster;
● Robot brain has the ability of self-regulation, which can quickly adapt to the aging and wear of robot body.
| Complexity | People with 3 years of experience | XingHuo’s robot | Efficiency improvement | |
| GPS antenna | 8 trimming sites | trim 1300 a day | trim 4000 a day | 3.1x |
| 5g filter | 24 trimming sites | trim 160 a day | trim 300 a day | 1.9x |
| 4G duplexer | 24 trimming sites | trim 20 a day | trim 40 a day | 2x |
Intelligent trimming
XingHuo technology released the world’s first 5g filter intelligent trimming robot, helping 5g industry development
In the 5g era, the number of channels of base station increases greatly, so the demand for RF components also increases. According to yole’s estimation, the global RF component market will increase from 105 billion yuan in 2017 to 245 billion yuan in 2023, and the market share of the filter, which plays a key role in the 5g industry, will increase from 56 billion yuan to 157.5 billion yuan. It can be seen that the filter is the fastest growing RF front-end device in the next few years.
Market scale development trend of 5g RF front end module from 2017 to 2023 (source: yole)
Compared with the traditional metal cavity resonator, ceramic dielectric resonator filter has the characteristics of high Q, low insertion loss and good temperature drift, so it has become the mainstream filtering scheme in 5g communication field.
Nanjing XingHuo Technology Co., Ltd. is a high-tech enterprise established with the strong support of Tencent, Alibaba and Baidu. In order to help the 5g industry, XingHuo technology has launched the world’s first intelligent trimming robot for 5g filter based on artificial intelligence. It uses the newly upgraded artificial cloud platform to efficiently complete the trimming of 5g ceramic filter.
Technical challenges
In the production process of ceramic dielectric filter, the most manpower dependent link is trimming. For a long time, the trimming of ceramic dielectric filter is to remove the silver on the filter surface by hand-held grinding tools, monitor the performance change of filter in real time through network analyzer, and then trimming according to the performance change continuously.
In addition, in the mass production process of filter, affected by the material inconsistency, process error and other factors, each product must be manually trimmed repeatedly to achieve the performance of the filter, so the trimming work has high requirements for manual experience. It takes a long time to train experienced trimming personnel, which greatly hinders the large-scale production of products.
Trimming automation is a difficult problem in the industry. Automatic trimming is to fix the dielectric filter on the manipulator, get the performance of the filter through the vector network analyzer, transmit the data to the computer, analyze the data through the algorithm, and then give the control signal to drive the manipulator to remove the silver on the surface of the filter.
Taking the current mainstream 10 order ceramic filter as an example, its structure is very complex, with up to 40 trimming sites and large decision space. Most importantly, every decision must be accurate enough, because once the silver is polished off, it cannot be recovered. If the decision is wrong, the product will be scrapped, which requires the robot to have a smart brain, bright eyes and flexible hands.
The unique secret of XingHuo technology
In order to meet the challenge of 5g ceramic filter, XingHuo technology has launched an intelligent cloud platform, which uses the most advanced couple matrix technology to extract the parameters of the filter, and uses deep reinforcement learning technology to train the agent from scratch to control the movement of the robot, and has achieved good results. XingHuo technology establishes a whole set of virtual simulation environment for robot, uses 56 GPUs for collaborative training, iterates 12 million times, and achieves good convergence.
The agent obtains the environment information through the optical system, and can control the robot to trim any site in the filter.
When an agent processes an optical image, it encodes the image as a potential representation and then uses it as a state input instead of the image itself. The key is that in the process of image coding, the model can learn the compact representation of potential state, so as to achieve accurate target capture. For this reason, XingHuo technology introduces a whole set of simulation environment, which contains a potential variable model that explicitly represents the potential linear dynamics, and takes this model as the basis of the algorithm.
Advantages of XingHuo technology artificial intelligence cloud platform
➤ the trimming time of 6th order ceramic filter is less than 10s; the trimming time of 8th Order ceramic filter is less than 30s; the trimming time of 10th order ceramic filter is less than 60s.
➤ learn to trim independently without human experience
➤ automatic loading and unloading
About Nanjing XingHuo Technology
Nanjing XingHuo Technology Co., Ltd. is a high-tech enterprise established with the strong support of Tencent, Alibaba and Baidu, specializing in artificial intelligence design, integrated circuit key technology, system cloud platform, industry application research and industrialization. The company always adheres to independent research and development, and owns key intellectual property rights. Its research and development progress is equal to that of Google in the United States, which is at the world’s leading level.
