Toshiba's "SATLYS" analytics artificial intelligence (AI) is an industrial AI that collects the knowledge Toshiba has accrued through its many years of manufacturing experience. Internet of Things (IoT) big data can be used to create a variety of business value and accelerate digital transformation. It leads the way to customer transformation and growth through improved productivity, more efficient business processes, and the creation of unprecedented business models. It digitizes the know-how of seasoned craftsmen and even leads to new discoveries that surpass those of experts. Let's look at the key advanced technologies used in SATLYS, the AI optimized for industrial applications.
What are the AI bottlenecks when using AI for industrial applications?
Creating AI that is truly useful for industrial applications is difficult even for global top class ICT vendors and software companies. Only companies with advanced AI technology and knowledge and expertise in various worksites can accomplish this. Toshiba's "SATLYS" analytics AI is an AI service that embodies the technologies and experience of the Toshiba Group, developed over the course of 140 years in the manufacturing field.
Many companies are trying to use the recognition and inference functions at which AI excels in their industrial and business workplaces. However, modern AI faces countless challenges that it must overcome in order to actually transform business and create new value.
For example, in worksites requiring immediate response, how can the tremendous amounts of data constantly being gathered by sensors and cameras, etc. be quickly and accurately analyzed? In situations in which abnormalities are rare, how can sufficient learning data be gathered? Can the basis of AI inferences be identified and explained? Going further, can AI that is useful to all customers truly be created for various business needs and management issues?
SATLYS provides answers to these questions, overcoming the limitations of existing AI to open a new frontier of industrial AI. It provides everything from classification and forecasting, cause inference, abnormality detection, and failure prediction to value chain optimization and automatic process control. It is a new analytics AI for the wide-ranging issues faced by industrial fields. This AI is backed by our highly polished, advanced technologies and know-how, developed through our many years of experience in a rich range of worksites.
Optimizing AI for industrial uses through advanced technologies and know-how
When it comes to AI, there is a tendency to focus on hot topics and algorithm capabilities. The fact that AI is often seen as a fear-inspiring technology that steals work instead of doing it on the behalf of people, and the focus on competition regarding superficial capabilities, may be reasons for this.
In developing SATLYS, however, we redefined AI as a technology meant purely to support people. Our goal was to optimize all of its functions for industrial applications as a grounded AI that will be of true use in business sites.
What was most helpful in achieving this were the diverse industrial field knowledge we have accrued through our past manufacturing and system integration work and advanced technologies such as deep learning developed through AI verification testing by ourselves, and through collaborations with customers and partners. Below are some of the key advanced technologies used in SATLYS, the culmination of our unique expertise.
(1) Large-scale image classification and big data analysis capable of handling tens of thousands of dimensions
Using AI in industrial fields requires recognition and inference accuracy and speed to be maximized. SATLYS contains AI platform technologies including high-speed, high-accuracy deep learning analysis. The innovative nature and dramatic results of these technologies have been proven through numerous verification experiments. Toshiba Memory Corporation's Yokkaichi plant, which manufactures semiconductor flash memory, is a typical example. The factory works with some of the most data in the world, processing two billion records per day. Every day 300,000 defect detection images taken with a scanning electron microscope are automatically classified. Conventional automatic classification technologies for defect detection images only detected 49% of defects, but deep learning has made it possible to automatically detect 83% of defects. Defect data from total inspections of wafers is used by the AI to automatically analyze the causes of defects. The analysis time taken per defect has been dramatically reduced, from an average of six hours to just two hours. SATLYS also provides big data analysis capabilities for analyzing large volumes of unstructured data. It has been proven to be highly effective in worksites which require rapid action in the form of real-time decision-making.
* The Toshiba Memory Corporation Yokkaichi Plant's case example is discussed in greater depth in #05.
(2) Achieving high-precision inference using limited learning data
When there is little learning data available, the precision of inference results falls significantly. This characteristic of deep learning presents a major challenge to the commercialization of industrial AI. SATLYS uses Generative Adversarial Networks (GAN) technology to automatically generate large amounts of learning data equivalent to actual data. This makes it possible to produce high-precision inferences even when there is little learning data (Fig. 1). This technology is used in industrial drone-based observation and inspection systems for power infrastructure. Based on the large amounts of generated learning data, the systems identify small abnormalities in power transmission line images, greatly improving inspection work efficiency.
* Automatic generation of learning data is introduced in DiGiTAL T-SOUL Vol.20 #02.
(3) Intuitive explanations through the visualization of abnormality causes
To tackle the problem of the bases of AI decisions becoming black boxes, SATLYS provides a groundbreaking function for visualizing which processes produce inference results. By visualizing AI -- clarifying what the AI is focusing on and on what it bases its decisions -- it supports environments in which developers and experts can provide intuitive explanations (Fig. 2).
Providing the AI solution as a professional service that tackles management issues
SATLYS also contains unique Toshiba technologies for automatically adjusting and optimizing the complex hyper-parameters that make up neural networks. We are constantly carrying out research with an eye towards autonomously evolving inference models when precision levels fall by monitoring the precision of AI inference results.
* Network optimization technologies are introduced in DiGiTAL T-SOUL Vol.20 #04.
This complete, polished AI solution is provided as a unified professional service that encompasses everything from digital consulting to system integration and operation support.
While Toshiba's "RECAIUS" communication AI business model standardizes various media intelligence technologies and provides them to an even greater number of customers, SATLYS precisely meets customers' varied but specific management issues. Through this, it leverages advanced technologies to create AI environments optimized to individual customers. It works together with customers to resolve issues that become evident in a wide range of fields, including not only manufacturing but also buildings and facilities, distribution and logistics, societal infrastructure, and energy (Fig. 3).
SATLYS use is already underway. Toshiba's AI is accelerating digital transformation in industrial fields, pioneering a new era.
* The corporate names, organization names, job titles and other names and titles appearing in this article are those as of January 2018.