Home > Technologies > From the Frontiers of R&D and Engineering > Research on data mining supported by wide-ranging business of Toshiba Group


From the Frontiers of R&D and Engineering

折原 良平
Research on data mining supported by wide-ranging business of Toshiba Group
Ryohei Orihara

As a consequence of population aging combined with a declining birthrate, the shrinking of Japan’s workforce has become a pressing issue. Many industries are under pressure to save labor through technological innovation and productivity improvement. One of the technologies expected to deliver a solution to this problem is data mining, or the process of analyzing a huge amount of data sets to discover meaningful patterns. The Toshiba Group has transformed many businesses by applying data mining technology to the accumulated data.

Ryohei Orihara

Data mining in the early days

I joined Toshiba Corporation in 1988 at the height of the economic bubble. Backed by ample funds, I devoted myself to a major research theme that explored how to support the thoughts and ideas of human beings. Take, for example, the food you have in your refrigerator. We human beings tend to be constrained by our experience and knowledge, becoming stuck in a menu rut when deciding what to prepare for a meal. However, artificial intelligence combined with a plethora of recipes available on the Internet can suggest all sorts of creative meals that you cannot think of on your own, based on the information on what is in your refrigerator. Artificial intelligence is conducive to cooking because a huge variety of recipes and organized knowledge about ingredients and cooking methods are readily available. Of course, I admit that there are areas where computers cannot beat human beings such as traditional Japanese cuisine with its delicate flavor and texture.

I also did research on predicting supermarket sales since the Toshiba Group manufactures and sells point-of-sale (POS) systems. Analysis of a huge amount of data on the products and their sales volumes from a POS system indicated that a certain poorly selling product was slightly too expensive. We tend to consider that it is hardly sensible to keep overpriced, slowly moving items on store shelves. On the contrary, data mining suggested that, without this product, the sales of a good seller would decline. The good seller was enjoying brisk sales because consumers found it reasonable when they compared its price with that of the expensive product. As demonstrated by this example, data mining can uncover hidden rules and causal relationships.
My parents ran a kimono shop. According to their experience, it was effective to recommend both low- and high-priced rolls of cloth at the same time. I was enthralled to see that data mining proved the human insight acquired through long experience.

I remember yet another episode that I experienced during the early days of data mining. It was around 2005. The Internet was already available, and many people were posting testimonials and reviews about the products they had bought on bulletin boards and other online venues. Our product planners and engineers had traditionally surveyed market demand and consumer perceptions, but most of these surveys were labor-intensive tasks. However, with the prevalence of the Internet, the amount of data to be analyzed increased dramatically. I realized there is a limit to labor-intensive surveys. To address this problem, we developed an Internet-based data analysis tool. The focus of this tool was on finding smoldering users’ grievances in Internet postings and identifying quality issues that invoked discontent without setting any explicit keywords. This is nothing special now, but, back in the mid-2000s, it was a groundbreaking approach. The Internet-based tool was proven useful in solving problems when it discovered important postings on the Internet that had previously escaped our attention. This is just one example that demonstrated how data mining improves work efficiency and overcomes labor shortages.

Some people balk at or do not trust new technology. However, data mining gradually gained wider uptake within the Toshiba Group through repeated successful applications.

Categorization of data analysis applications

Table 1. Categorization of data analysis applications
Selective data analysis is a type of analysis based on simple conditions whereas comprehensive data analysis refers to derivation of a definition of a concept based on data. The gaps between analysis results and human decisions (that have a direct or indirect impact) consist of another dimension of the classification.
Reference:Toshiba Review, VOL. 69 (2014)

Deployment at a fab — Conflict and success

The semiconductor fab at Toshiba’s Yokkaichi Operations recently undertook an initiative to improve productivity. In order to meet the ever-increasing market demand for higher quality and price reduction, it is always important to reduce manufacturing defect rates. The Yokkaichi fab was gathering more than two billion data items every day from automatic transport systems, test systems, and product management systems supporting tens of thousands of processes. These data items were visually represented as graphs and tables, which were analyzed by experienced staff.
However, as the manufacturing processes became increasingly complicated, the data analysis task was becoming more time-consuming. In these circumstances, Yokkaichi Operations decided to try out data mining.
Since fab managers had yet to recognize the effectiveness of data mining, the initiative began at the grassroots. It was therefore difficult for the researchers and fab engineers to work closely together. Indeed, after making a request, it took two weeks until we data mining researchers received the data we wanted. Data mining would not have delivered any benefit if our communications had remained so slow.

Feedback from fab engineers is essential in order for us to understand how we should interpret data. Haphazard data analysis would not get us anywhere. To break the deadlock, the researchers stayed at the fab for a few weeks to learn about the fab processes and improve communication with the fab engineers. In addition, the deployment of BigReport, an integrated database system, at the fab facilitated data-sharing between the researchers and fab engineers. Our persistent efforts led to the development of a comprehensive big-data-based monitoring system for yield analysis. While it took an average of six hours to manually analyze a single defect, the system reduced the analysis time by two-thirds to two hours.

This would not have been achieved with data mining technology alone. Since we researchers acquired in-depth knowledge of products at the fab and established a relationship of trust with the fab engineers, we were able to deliver results in both R&D and manufacturing. I believe activities like this are unique to Toshiba.

Expansion of data mining

The more data you have, the greater the benefit data mining delivers.

It was in the 2000s when the term “big data” appeared. As described above, we recognized the importance of collecting and analyzing huge amounts of data before this term was coined. As we operate in a wide range of business, we have the optimal environment for data mining research. Our applications of data mining encompass diverse areas, including aging diagnosis of social infrastructure using data on social media; lifetime analysis of elevator components using maintenance logs; improvement of power generation performance of photovoltaic (PV) plants, taking various plant and weather conditions into account; patents; and stock prices.

Data mining is also applied to the results of physical checkups. This application uses numerical data from physical checkups conducted in the past few years so as to provide health assistance. It was realized by utilizing a large amount of accumulated data with the cooperation of Toshiba’s employees. This was possible due to the large employee population of the Toshiba Group.

Data analysis technology of Toshiba Group

Figure 1. Data analysis technology of Toshiba Group
The Toshiba Group has been developing advanced analysis techniques and databases using big data acquired from diverse business areas ranging from social infrastructure facilities and systems to consumer electronics.
Reference: Toshiba Review, VOL. 69 (2014)

Ryohei Orihara

Next step

It is indeed a pleasure to write a computer program and discover something new derived from that program. I have learned a lot as I apply data analysis technology from various perspectives.

I also enjoy the opportunities for activities in the wider research community.
I am a member of the Japanese Society for Artificial Intelligence (JSAI) and the Information Processing Society of Japan (IPSJ). I also participate in study groups to meet new people and I teach at a university.
I find these activities inspiring, and, for data mining, it is important to encounter new perspectives. Dialogs with researchers in the wider community and with university students also help broaden my views while deepening my knowledge.

I am privileged to be able to do research in cooperation with various divisions of the Toshiba Group operating in a wide range of businesses while at the same time drawing inspiration from the wider research community. I would like to take advantage of this favorable environment to look at social issues from different angles with a view to resolving them.

Ryohei Orihara

Ryohei Orihara, Ph.D.

Data mining, big-data analytics
Corporate Research & Development Center, Chief Research Scientist
Registered patents:
81 (Japanese patents)
Academic societies:
The Japanese Society for Artificial Intelligence (JSAI), opens in a separate window
Vice Chairman
Information Processing Society of Japan (IPSJ), opens in a separate window
2017 2016 JSAI Field Innovation Award
2013 IPSJ Activity Contribution Award
2011 JSAI Distinguished Service Award

* Went live on March 30, 2018. This webpage is based on the latest information available as of March 30, 2018.

opens in a separate window Link with this mark opens in a separate window