Toshiba Analytics AI Underpinning Acceleration of Digital Transformation Future Outlook for AI in Reiwa Era and Toshiba’s Approach YAMAMOTO Hiroshi Toshiba Analytics AI Technologies for Creation of New Value through CPS KOTO Shinichiro / NISHIKAWA Takeichiro / KAWAMURA Takuya / YOSHIKAWA Takashi In the manufacturing and social infrastructure fields, the effects on production stability and maintenance of quality arising from the dependence on implicit knowledge-based operations by skilled workers and the complexity of systems and operations have become a matter of concern in recent years accompanying the shift to an aging society with fewer children in Japan. The recent increase in traffic accidents caused by elderly drivers has also become a social issue. In order to address these and various other social issues, the Toshiba Group is actively promoting research and development aimed at applying artificial intelligence (AI) to motifs in the real world. We have set the goal of becoming a cyber-physical systems (CPS) technology company in order to offer new value to customers by combining component technologies in the physical space and technologies in the cyber space based on AI and the Internet of Things (IoT). These AI technologies cultivated through our development experience accumulated over more than half a century are expected to be newly deployed in a broad range of applications as key technologies for CPS. Fault Diagnosis Method Based on Machine Learning using Time-Series Sensor Data UENO Ken / YAMAGUCHI Akihiro / MAYA Shigeru The progress of Internet of Things (IoT) technologies has expedited the collection of time-series sensor data from various equipment in social infrastructure facilities and manufacturing sites. Particularly in the case of facility monitoring systems, attention is being increasingly focused on the effective utilization of large volumes of time-series data for anomaly detection and prediction. In this context, Toshiba Corporation has developed a fault diagnosis method incorporating the following methods based on machine learning: (1) anomaly detection to warn of the risk of changes from normality to abnormality with certainty, (2) anomaly detection to offer grounds for abnormality determination by learning normal waveform patterns of time-series data, and (3) anomaly prediction to follow numerous state variations by continuously learning the waveform patterns of time-series data. We are now carrying out the verification of elemental performance aimed at the practical application of these methods to the operation and maintenance of a wide variety of facilities. Automated Defect Classification System for Semiconductor Manufacturing Processes Using Deep Learning IMOTO Kazunori / NAKAI Tomohiro In semiconductor manufacturing processes, defect classification and monitoring using scanning electron microscope (SEM) images acquired in the wafer inspection process are essential for early detection of process abnormalities and improvement of production yield. However, conventional defect classification systems depend on visual confirmation due to a lack of classification accuracy in these complex manufacturing processes. The Toshiba Group has developed a defect classification system that can automatically classify each defect in SEM images with high classification accuracy. To reduce deterioration of classification performance when reliable labeled data are lacking, this system utilizes two types of deep learning methods: a weakly supervised learning method and a transfer learning method. We have conducted evaluation experiments using actual data at our semiconductor manufacturing factory and confirmed that these methods achieve a higher classification accuracy rate than conventional methods as well as a high classification accuracy rate even in the case of a limited number of reliable labeled data. We have also confirmed the effectiveness of our system in terms of increasing the learning speed necessary for practical application. Insulation Deterioration Diagnosis Technique for Switchgears Using Machine Learning BANNO Kozo / NAKAMURA Yusuke Switchgears are a key type of power supply equipment that contain circuit breakers in a metal casing. If a switchgear’s insulation performance deteriorates due to aging, there is a possibility that failure of the switchgear might occur, resulting in a serious incident in the electric power system including a blackout. Demand has therefore been increasing in recent years for insulation deterioration diagnosis techniques for various types of switchgears, in order to eliminate the dependence on inspection work currently being carried out by a limited number of skilled engineers. With this as a background, Toshiba Infrastructure Systems & Solutions Corporation is developing an insulation deterioration diagnosis technique for switchgears focusing on the phenomenon of partial discharge, which is a precursor of insulation deterioration in solid insulators. This technique makes it possible to perform a diagnosis based on the inference of locations and types of defects from the relationship between partial discharge signals and defects caused by insulation deterioration, by means of machine learning. We have confirmed the effectiveness of this technique through experiments using different types of samples. High-Precision Electricity Demand Forecasting Method Using Weather Prediction Data and Deep Learning SHIN Hiromasa / SHIGA Yoshiaki / ICHIKAWA Ryoichi With the proliferation of renewable energy systems in Japan, which are affected by weather conditions, it has become necessary to make further efforts to maintain the electricity supply and demand balance as well as the amount of electricity generated. As the profits of electricity retailers are directly linked to the accuracy of demand prediction under the current electricity system, there is a strong requirement for highly precise forecasts of electricity demand. The Toshiba Group has developed a high-precision electricity demand forecasting method using numerical weather prediction data and deep learning. This method provides customers with information on electricity demand through the following processes: (1) creation of meshed weather prediction data for all areas of the country based on numerical prediction data distributed by the U.S. National Centers for Environmental Prediction (NCEP), (2) feature extraction from a small amount of information by means of a sparse modeling method, and (3) reduction of prediction errors by means of an ensemble learning method. The use of this method has achieved a reduction in prediction errors of approximately 1% in calculations using the data of the targeted areas for the preceding day compared with the conventional forecasting methods. We are aiming to apply our new method to electricity management systems for electricity retailers and virtual power plant (VPP) services. Low Power Consumption Deep Convolutional Neural Network Accelerator for ADAS ISHIGAKI Yutaro / TANABE Ken / TANABE Yasuki Attention has been increasingly focused on the introduction of artificial intelligence (AI) technologies, particularly recognition and classification technologies applying deep convolutional neural networks (DCNNs), that can play a key role in realizing the safety of automobile driving operations using advanced driver assistance systems (ADAS). However, as DCNN processing involves a huge number of calculations and massive volumes of data, it is difficult to apply it to automotive systems with limited power consumption. To overcome this problem, Toshiba Electronic Devices & Storage Corporation has developed a real-time hardware accelerator (HWA) with low power consumption for automotive applications. This HWA makes it possible to efficiently implement DCNN processing by achieving reductions in the number of memory accesses and the volume of data. Accurate Real-Time Video Analysis System to Recognize Road Relay Race Teams in Live Images Using Deep Learning YAMAJI Yuto / KOBAYASHI Daisuke / SATO Makoto The field of sports broadcasting has recently seen an accelerating movement toward the introduction of functions utilizing information and communication technologies (ICTs) to grasp the situation of sports events and automatically create computer graphics (CG) superimposed on video images. With the aim of improving the efficiency of live TV production of road relay races, Toshiba Corporation has developed a real-time video analysis system that automatically recognizes the team of each runner. In a road relay race, it is difficult to recognize runners’ teams because of the frequent overlapping of runners and constantly changing outdoor lighting conditions. By using a deep learning model, our newly developed system makes it possible to recognize runners’ teams based on their uniforms, which can be readily tracked even when runners overlap, and to identify each team with a high degree of accuracy based on learned uniform logos and colors under various outdoor lighting conditions. A demonstration experiment on this system at an actual road relay race confirmed that it achieves a practical team recognition accuracy of 98.1% in live broadcasting and contributes to the reduction of conventional manual confirmation work. Learning Methods for Realization of Optical Character Recognition with High Accuracy Using AI FURUHATA Akio / TANAKA Ryohei / OSADA Kunio Attention has been focused in recent years on robotic process automation (RPA) using software robots amidst the ongoing improvement of routine, repetitive business operations in Japanese companies. As part of these efforts, the introduction of optical character recognition using artificial intelligence (AI-OCR), which makes it possible to recognize various handwritten characters on a wide variety of business forms with a high degree of accuracy, is expected to improve the efficiency of paper-dependent business processes. However, the higher data collection and teaching costs of AI-OCR, due to the need for a large number of different character samples including a vast number of handwritten kanji characters in order to improve recognition accuracy, are a serious issue. To rectify this situation, Toshiba Digital Solutions Corporation is engaged in the development of the following methods for AI-OCR to improve character recognition accuracy while suppressing increases in costs: (1) a data augmentation method that can generate a variety of character data based on a small number of actual handwritten data and (2) a semi-supervised learning method using virtual adversarial training to recognize character strings. We have confirmed the effectiveness of these AI-based methods through verification tests. Technique to Provide Users with Grounds for Inference Using SATLYS Toshiba Analytics AI Services TAKAHASHI Shintaro / KATO Toshiyuki / NISHIZAWA Minoru Technologies related to artificial intelligence (AI) have begun to be introduced in the manufacturing and social infrastructure fields in recent years in order to overcome dependence on implicit knowledge-based operations and reduce operating costs. This has led to the need for a technique to provide users with grounds for the results inferred by AI so as to implement appropriate measures. Toshiba Digital Solutions Corporation has developed and is supplying SATLYS Toshiba Analytics AI services, which incorporate high-accuracy analysis classification, prediction, factor estimation, anomaly detection, fault sign detection, and activity estimation functions, based on the knowledge that it has cultivated through Toshiba’s experience in manufacturing. We have also developed a technique to provide users with grounds for the results inferred by SATLYS. This technique offers useful information for finding solutions, taking advantage of the functions of SATLYS including a defect factor analysis function capable of detecting defects in die-cast products and a function to display pricing factors capable of automatically scrutinizing estimates from suppliers in the procurement process. SATLYSKATA Cloud Services Offering Sophisticated Analytics without AI Experts SAWADA Shoichi / YAMAMOTO Junichi / KOBAYASHI Kenji SATLYS Toshiba Analytics AI services, a suite of artificial intelligence (AI) analytics solutions for industrial markets covering all areas from data analysis through system construction to cloud services, are contributing to the resolution of customers’ issues and the realization of digital transformation. In order to offer SATLYS services to more customers, Toshiba Digital Solutions Corporation has developed SATLYSKATA AI analytics services customized for specific purposes. SATLYSKATA services are cloud services in which proven analytics models are standardized, incorporating an analysis execution environment and application programming interface (API). The first release comprises two services: SATLYSKATA Maintenance Parts Inventory Optimization and SATLYSKATA Work Activity Estimation. Demonstration experiments on these services have verified that SATLYSKATA Maintenance Parts Inventory Optimization achieves an approximately 30% reduction in surplus inventory parts related to maintenance and that SATLYSKATA Work Activity Estimation achieves an approximately 15% reduction in work time. We are expanding the lineup of SATLYSKATA services with enhanced functions. |