In April 2018, Taipei – ArTIficial Intelligence (AI) is undoubtedly the hottest topic in the technology industry in the past year or two. In addition to the huge investment by technology giants, financial services providers have also introduced artificial intelligence. Strong interest. The manufacturing industry's attention to AI technology is also not a problem, and in the case that relevant key technologies are gradually in place, actual import actions have begun. Advantech, which advocates smart manufacturing and spares no effort, not only provides corresponding advanced solutions for all walks of life, but also gradually introduces artificial intelligence elements in its own production online. For example, the status detection/diagnosis of the machine equipment, the use of raw materials/energy, and even the quality control process of the products have gradually introduced artificial intelligence. Arm's Silicon Intelligence (IP) and SoC and Xilinx's Field Programmable Gate Array (FPGA) technology are two of the best tools for Advantech to promote production line AI. Advantech IoT.SENSE interviewed Advantech's technical director Yang Ruixiang and the director of the plant, Lin Dongjie, to discuss Advantech's solutions under the innovation wave of AI, IoT, and intelligent manufacturing. The following is a summary of the interview: Yang Ruixiang, director of Advantech's technology, said that artificial intelligence is not a new topic in the field of academic research. The main reason why the public and the industry have been widely watched in the past two or three years is that its evolution rate is too amazing and can be created. A huge commercial value is no longer just a topic of academic research. In addition to targeting specific areas, AI technology is also pursuing higher versatility. Deepmind's latest chess program has removed Go (Go), called Alpha. Because the program also knows that Japan will play chess and other chess classes, and successively defeat other world-class special chess programs. This is undoubtedly an important milestone in the development of artificial intelligence. In the case of rapid evolution and huge commercial value, artificial intelligence has become the most eye-catching technology issue, which is not surprising. But the discussion goes back to the discussion. How to introduce artificial intelligence in all walks of life to realize the intelligence of the industry, there are still many details to overcome. In the manufacturing industry, regardless of the final product, the manufacturing industry can't escape the four elements of "people, machines, materials, and law." People refer to employees, and machines refer to various tool machines. Materials refer to various raw materials and energy sources. The law is the process method. Since the industrial revolution, no matter how the products of the manufacturing industry evolved, these four elements cannot be separated. How to optimally manage these four elements is a topic that manufacturers face every day. Yang Ruixiang analyzed that the introduction of artificial intelligence, the most important four KPIs, is to show the optimization and improvement of the artificial materials method. In terms of people, how to turn the experience of the master into quantifiable parameters, and then copy and spread the experience of human beings is an important goal of importing AI. However, to achieve the above four optimizations, the most important thing is the degree of understanding of the AI ​​and the quality of the collected data sets. First of all, manufacturers must have a correct understanding of AI, know what problems AI is suitable for, and what restrictions apply. Secondly, the training results of the AI ​​inference model, in addition to the design of the model itself, the quality of the training data is also very important. If the model is trained with quality-quality data, the results of the AI ​​inference will fall from the reality. Finally, the organizational culture has to be adjusted. Before the introduction of AI, all the decision makers on the production line were people, relying on past experience; after the introduction of AI, although the final decision is still human, but no longer only subjective feelings or experience to judge, and It is a relatively objective statistical science. The trust relationship between people and machines takes a while to improve. Of course, AI itself must continue to evolve, improving the reliability and accuracy of its predictions. Yang Ruixiang further explained that artificial intelligence can be divided into two parts: training and inference. For production site applications, most of the trained models are used to perform various inference applications. Model training is not directly performed at the edge because model training requires powerful computing power and a large data set, which is more suitable for data center or On the cloud. Also because of the low demand for computing power, there are many off-the-shelf processor solutions available on the market. For example, x86 CPUs, GPUs, and Arm-based SoC processors can perform related computing tasks. The only difference is that Cost, power consumption and heat dissipation can meet the specifications of field devices. From a technical point of view, GPU is the most suitable processor architecture for model training. It is certainly more than enough to perform model inference tasks, but the cost of GPU and the heat dissipation problem that power consumption are such. The biggest limitation of the processor on edge nodes or field device applications. The x86 CPU also has very powerful computing power, but because its architecture is designed to meet a variety of computing/control applications, it is not as efficient as the GPU when executing AI algorithms. Yang Ruixiang analyzed that this issue is related to the nature of AI. AI usually only processes a large amount of data with a few instructions or even a single instruction. For example, the Deep Learning and Convolutional Neural Network (CNN) is a matrix operation from a mathematical point of view, which is very similar to a drawing operation, so the GPU naturally has an inherent advantage in this respect. The x86 CPU is longer than the multi-instruction stream multi-data stream (MulTIple InstrucTIon, Multiple Data, MIMD) computing environment, but when the data volume is too large, it must be pulled high frequency, or multi-core and more Thread architecture to deal with. The Arm Processor with Reduced Instruction Set (RISC) has innate features between the GPU and the x86 CPU. In addition, the performance of the single instruction, multiple data (SIMD) of the Arm processor has been strengthened in recent years. Therefore, it is more convenient when performing AI operations. Although model training is currently required with the Arm processor, it is not comparable to the GPU in terms of efficiency, but it is the most balanced solution for power consumption, cost, and performance when performing inference tasks. Yang Ruixiang revealed that in recent years, Advantech has worked closely with Anmou and has a certain grasp of the product development blueprint of Anmou. Future security will introduce a more specialized, more efficient processor core and peripheral IP for AI computing needs. This will be a great help for promoting the popularity of edge computing and AI applications. Advantech will also continue to maintain a close cooperative partnership with the company. Tightly grasping the four elements of the artificial materials method, Advantech has begun to use the Arm architecture SoC and Xilinx FPGA module as the hardware foundation, and gradually introduce artificial intelligence in its own production online. Advantech General Manager Lin Dongjie said that Advantech is currently importing AI in production and has entered the stage of using AI to assist in the interpretation of the original data. In the era of industrial Internet of Things, not only the production of online machines will generate a large amount of data, but also the infrastructure of the plant will generate considerable data volume. It is not time-sensitive and has limited benefits to use humans to interpret these data and analyze the meaning behind it. Finally, because the environment in which Advantech is located is typically a small number of diverse, single-production models, which are very different from the general consumer product specifications and mass production, the management of the production line is relatively complicated. This is also one of the pain points that Advantech hopes to solve when introducing artificial intelligence. Lin Dongjie said that due to technical limitations, the ultimate goal of fully interpreting raw data by the system is not yet available, but this is the direction of Advantech's future efforts. More specifically, in the future, Advantech's intelligent manufacturing hopes to achieve three major goals: First, the modernization of production equipment, hope that all machine equipment can support Industry 4.0; Second, to achieve the interface between data acquisition and software, mainly Data is connected to systems such as Manufacturing Execution System (MES) and Product Lifecycle Management (PLM). Third, machine vision and deep learning are further expanded and applied in quality control. In response to the first point, Lin Dongjie did not say that the upgrade and transformation of the existing machine is usually not cheap, especially if the original factory needs support or authorization, and cannot change it by himself. However, in some cases, the existing machine has been able to obtain sufficient parameter data through the data acquisition module developed by Advantech itself. As for the expansion of machine vision and deep learning, Advantech is currently working with the China Academy of Sciences to develop a machine vision system that can detect a variety of different products. In fact, Advantech has been using optical automatic detection (AOI) for a long time, but the existing AOI system is only suitable for the detection of fine components on the motherboard and circuit board, and is not suitable for detecting finished products or larger zeros. part. On the other hand, the small and diverse characteristics of Advantech products also make the current machine vision solutions on the market to be used in Advantech's production line, encountering considerable difficulties. At present, most of the machine vision solutions on the market are designed for the inspection needs of a large number of products, but Advantech's demand is a machine vision inspection solution that can automatically adapt to various product types. Therefore, Advantech decided to work with the China Academy of Sciences to develop a customized deep learning algorithm to make the machine vision system smarter to adapt to different types of products. Machine vision is the stage for the FPGA module to show its talents. It is also one of the projects that Advantech's FPGA application development team has made concrete achievements. Through the FPGA module, Advantech is free to decide which image recognition links need hardware acceleration to improve the performance of the visual inspection system. Yang Ruixiang pointed out that in addition to the CPU and GPU, using a dedicated hardware acceleration chip to improve the performance of the AI ​​system is theoretically a feasible way. However, the current AI algorithm is still evolving rapidly. If an ASIC is adopted, it is likely that it will not catch up with the pace of technological development. FPGA is a compromise between performance and flexibility. The structure of the computing unit can be customized to meet the needs of specific algorithm acceleration. Because of the programmability, when the algorithm needs to be modified or updated, there is no need to reopen a chip. Just modify the design code. Therefore, at this stage, FPGA is one of the ideal solutions for AI algorithm acceleration. Advantech also has a mature FPGA application development team, and will continue to invest in this technology in the future. Advantech Embedded DTOS FPGA Capability Plastic Speakers,Plastic Speaker With Labtop,Plastic Abs Speaker,Portable Plastic Speaker Comcn Electronics Limited , http://www.comencnspeaker.com
