MyAdvantech Registration

MyAdvantech is a personalized portal for Advantech customers. By becoming an Advantech member, you can receive latest product news, webinar invitations and special eStore offers.

Sign up today to get 24/7 quick access to your account information.

Powered by Arm/FPGA, Advantech Enters AI in Manufacturing

4/9/2018

In the past two years, artificial intelligence (AI) has been one of the hottest topics in technology industries. Besides technological giants who have pioneered this new trend, many players in the manufacturing industry have taken action to realize AI applications, and such companies have had a particular advantage in this regard because many of the required technologies are already in place.   

As an advocate of intelligent manufacturing who has been providing advanced intelligent solutions for various industries, Advantech has also begun introducing AI elements (e.g., intelligent monitoring of machine status, raw material/energy usage, and QA/QC process) into its own production lines. Two major contributors for Advantech in adopting AI to modernize its production lines are intellectual property (IP) core technology from Arm and field-programmable gate array (FPGA) technology from Xilinx.   

Advantech’s IoT.SENSE staff interviewed Advantech Chief Technology Officer Allan Yang and Associate Vice President Jamie Lin to share their views on and strategies for innovative trends driven by AI, IoT, and smart production.  

Advantech CTO Allan Yang


Advantech AVP Jamie Lin

Amazing AI Development Pace and Huge Business Potential 

Mr. Yang noted that although AI is not a new topic in scientific research, it had not gained much attention from the general public or industries until the past two or three years. The main reason for the shift in attention is the pace of recent developments, which has created huge business potential in areas outside of academia.

A major milestone that has kindled the AI rush was when AlphaGo, a computer program developed by the Google subsidiary DeepMind, whitewashed South Korean professional Go player Lee Sedol in March 2016. Despite this achievement, DeepMind has continued to advance its developments with AlphaGo Zero, a self-trained version of AlphaGo that learns by playing against itself. This newer version has thrashed older versions of AlphaGo that have utilized thousands of human amateur and professional games of Go for training. This amazing breakthrough in AI has demonstrated the potential to go beyond what human knowledge has accumulated over thousands of years.

Developments in AI are currently aimed at higher versatility or multi-usability rather than a single focus on a dedicated area. DeepMind has renamed its chess program to AlphaZero, because the program can now also a variety of chess-type games, including Japanese Shogi. This new system has repeatedly outperformed other top computer chess programs in the world, and this is undoubtedly an important milestone for AI moving toward versatility.

In fact, the commercial value of AI had manifested itself long before the appearance of AlphaGo. For example, the use of YouTube’s video recommendation system and Amazon’s online shopping store are the direct result of deep learning, a technology branch of AI. Considering the fast evolution and potential commercial value of AI, it is unsurprising why AI is currently the hottest technology. However, talk is one thing and practice is another matter entirely. There are many details and hurdles that need to be overcome to put AI into practice in specific application scenarios in different industries.

Arm Architecture is Suitable for AI Inference Processing

Mr. Yang explained that AI can be roughly divided into two parts: model training and inference. Most field applications in production lines use already-trained models in implementing inference algorithms rather than directly training models at the edge; this is primarily because the training of AI models requires considerable computing resource and a massive data pool, the implementation of which is traditionally considered to be more for the cloud. By contrast, AI inference processing requires less computing power and can be implemented by many processing solutions that are available in the market, such as x86 CPU, GPU, and Arm-based SoC architectures.

From a technical perspective, GPUs are currently the most suitable architecture for training models, and are sufficient for inferencing implementations. However, GPUs have drawbacks in cost, power consumption, and heat ventilation, and these problems have confined the deployment of these processors in edge nodes or field applications. The x86 CPU architecture is offers outstanding processing performance, but this architecture was designed mainly for computational/control applications. When an x86 CPU is employed for implementing AI algorithms, however, it is comparably less efficient than a GPU.

Mr. Yang pointed out that the reason for this discrepancy is related to the nature of AI, which usually uses few instructions or even a single instruction in processing large amounts of data. From a mathematical point of view, the processing of deep learning and convolutional neutral networks is actually data-parallel array computation, which is very similar to graphical processing. Accordingly, GPUs have an innate advantage in this regard. By comparison, x86 CPUs outperform GPUs in dealing with multiple instruction, multiple data streams; however, when the data volume is sufficiently high, a higher clock rate or multiple cores or multi-threading is required in order to respond to the increased workload.

Arm processors based on the reduced instruction set computing architecture are the middle ground between GPUs and x86 CPUs.  With continual improvements in single instruction, multiple data processing in recent years, Arm processors have become increasingly more suitable for AI computation. Although Arm processors are still less efficient for training models than are GPUs, they give the best balance among power consumption, cost, and efficiency when implementing inference tasks.

Mr. Yang revealed that Advantech has developed a close relationship with Arm Holdings, and the company understands the product roadmap of Arm processors. For the future, Arm will target the needs of AI processing and launch more specialized and more efficient IP cores and peripherals, which will be a great help in promoting edge computing and AI applications.  Advantech will continue to maintain its close partnership with Arm.

AI Reaches the Production Floor with Magic Edge Computing

Committed to the idea of improving the four elements of manufacturing, Advantech has begun introducing AI into its own production lines, using Arm-based SoC and Xilinx FPGA modules as a hardware platform.

Advantech Associate Vice President Jamie Lin, who has the responsibility of managing all Advantech facilities, stated that the application of AI in Advantech production lines is currently being utilized to assist reading and judging data pertaining to raw materials. In the age of IIoT, data are generated at both the production line and infrastructure level to be accumulated into a large data pool. Using human brain power to read and judge such a sizeable volume of data in order to extract something significant and meaningful would be impractical because of the amount of time and effort required to do so, thus rendering such an approach ineffective and inefficient. Moreover, because humans cannot avoid bias and subjective views, it is likely that different persons would produce different judgments, which would be seen most often at the QA/QC stations.

It is important to note that Advantech's production differs from typical consumer product manufacturers, since the company operates under a high-mix/low-volume production model. Thus, the management of its production lines is more complex. This is one of the most challenging points for Advantech to overcome when introducing AI into its manufacturing processes.

Advantech’s AI production lines can now automatically collect parameter data on job orders, yield rates, machine operational status, component failure alerts, and so on. Even for old machines that were in use before the implementation of AI, Advantech can use its own solution to acquire parameter data such as temperature, shock/ vibration, voltage, and current, and so on, which the AI system can read and evaluate. However, according to Mr. Lin, it remains difficult to implement 100% AI judgment on all raw data due to technical constraints, but Advantech will continue to endeavor to reach that goal.

Mr. Lin revealed three major goals for Advantech’s intelligent manufacturing: 1) to modernize its production facilities, which involves enabling all machines and equipment to support Industry 4.0 implementation; 2) to realize the interfacing of data acquisition and software execution and import the data into its manufacturing execution system, product lifecycle management and other corporate IT systems; and 3) to expand the application of machine vision and deep learning to all nodes and links for QA/QC.

For the first goal, Mr. Lin acknowledged that it is costly to upgrade or retrofit existing machines and equipment, especially those that require support or licensing from the original equipment manufacturer. However, under certain circumstances, these machines and equipment may be equipped with add-on data acquisition modules developed by Advantech, and these can be installed with non-intrusive methods, thus providing a means to acquire sufficient parameter data.

As for expanding machine vision and deep learning applications, Advantech has cooperated with Taiwan’s Academia Sinica in developing a well-rounded machine vision system that can be universally employed to inspect all types of products. In fact, Advantech has employed automated optical inspection for some time, but this system can be applied only to the inspection of host boards and miniature components on electronic circuit boards instead of in final products or larger components.

Besides, with Advantech’s high-mix/low-volume production, there have been difficulties in adopting machine vision solutions in the company’s production lines. Currently, most machine vision solutions in the market have been designed to inspect products in mass production, but Advantech requires a versatile solution that can be applied to all types of products. This is why the company has decided to work with the Academia Sinica to develop customized deep learning algorithms so that the machine vision system can adapt to different product types more intelligently.   

Reconfigurable Acceleration Stack

FPGA Modules to Help Accelerate Machine Vision Processing 

FPGA technology may be the answer to Advantech’s problem. The Advantech FPGA team has developed an FPGA module for machine vision system in order to solve the pending problems. With this module, Advantech can flexibly decide which parts of image recognition processing should be accelerated by hardware depending on the application, thus enhancing the operational performance of the vision inspection system.

Mr. Yang noted that, in addition to CPUs and GPUz, a specialized hardware acceleration chip could theoretically be an option to improve the performance of AI systems. However, considering the rapid innovation pace of AI algorithms, using application-specific integrated circuit chips might fall behind technological developments. FPGAs are a compromise between performance and flexibility, and they allow for customization of computing kernels in order to meet the acceleration requirements of specific algorithms. The programmability of FPGAs also allows the chip to adapt to revised AI algorithms simply by reprogramming without the need for fabricating a new chip.

Therefore, at the present stage, FPGAs are one of the most ideal solutions for accelerating the processing of AI algorithms. Advantech has prepared a well-established in-house R&D team to explore FPGAs and will continue to invest on this technology.


Advantech Embedded DTOS FPGA Capability