Toshiba Develops an AI Technology for Automatically Generating Physical Models with Excellent Interpretability from Time-Series Device Data, Facilitating the Detection of Device Abnormalities and Their Cause

-Determining not only whether something is different but also why contributes to predictive maintenance of social infrastructure equipment-

5 November, 2021
Toshiba Corporation

Overview

TOKYO─Toshiba Corporation (TOKYO: 6502) has developed the AI technology that automatically generates physical models(*1) for detecting equipment abnormalities and determining the causes of those abnormalities in social infrastructure equipment. The AI automatically generates a physical model that describes the states and motion of a target device from measured time-series data. In addition, the AI shows not only abnormalities in equipment but also the cause of the occurrence of abnormalities based on physical phenomena, which has been difficult in the past. For example, the model made by the AI might indicate that an increase in equipment temperature is due to clogging dust, making improvement measures easier to formulate. The AI responds to complicated phenomena such as sudden changes in temperature and airflow, and can be applied to the maintenance of infrastructure equipment with complicated mechanism of abnormalities occurrence that is abnormalities with complicated occurrence mechanisms. It will improve reliability of facilities and contribute to resilience in social infrastructure. Toshiba will present the details of this technology on November 5 at IMECE 2021, which will be held online.

Development Background

Predictive maintenance, in which equipment deterioration and defects are detected in advance for management in an optimal state, is an increasingly important factor in the continued safe and secure use of products and systems. The market for predictive maintenance solutions is in the growth stage, reaching about $6.9 billion in 2021, with projections for rapid growth at a compound annual growth rate of 31% to about $28.2 billion worldwide in 2026(*2).
As an infrastructure services company, Toshiba is working to reduce interruptions and downtime due to failures in equipment that support social infrastructure, and to establish highly accurate predictive maintenance technologies that minimize maintenance costs. Anomaly detection and suggestions for solutions are indispensable aspects of predictive maintenance. However, many infrastructure devices generate abnormalities via complicated mechanisms, and an outstanding problem is the difficulty of formulating improvement measures by using conventional anomaly detection technologies that only show differences from the normal state. Taking appropriate measures and reducing maintenance costs will require technologies that can explain not only what is different from normal, but the causes of such differences.

Features of the Technology

Toshiba thus developed an AI technology in addition to detecting anomalies in equipment, that can use a physical model with excellent interpretability to explain why anomalies occur. This technology automatically generates a physical model that describes the state and operations of a device from time-series data measured from that device. In the automatically generated physical models, correlations between data items are represented as a network, and relations between items are represented as a combination of functions from physics and engineering. Candidate functions, which are stored in a database, are based on mechanical engineering knowledge that Toshiba has cultivated over many years, applying the handling of complex phenomena. Furthermore, conventional AI technology is problematic in that it is difficult to efficiently combine huge numbers of functions without changing their physical meaning. Toshiba therefore developed a new AI technology that combines a new sparse estimation algorithm that can correctly consider the degree of physical influence of a function, a spatial search algorithm for efficiently selecting candidate functions, and a data expansion algorithm that enables highly accurate prediction. Toshiba succeeded in development of the AI for automatically making physical models with excellent interpretability.

Figure 1: Overview of the physical model generation.

This technology eliminates the need for data such as the dimensions of a device and the physical properties of its parts, which have been required to generate conventional physical models, and it is advantageous in that it generates physical models using only measurement data from sensors, allowing regular model updates during product or system operations. Analyzing changes in updated physical models allows detection of signs of product or system abnormalities and identification of their causes.

Figure 2: Using the generated physical model to explain why a problem has occurred.

Applying the AI to temperature predictions, which is an important part of abnormality detection in power modules(*3), automatic generation of a physical model confirmed that a heat transfer method by which heat is transferred from a heat-generating chip to a cooler and radiated from the cooler by a cooling fan is correctly selected.
The generated physical model predicted temperatures with high accuracy (average error less than 1°C), and it was possible to realize real-time predictive maintenance instead of performing a detailed numeric simulation, which would take thousands to tens of thousands of times more calculation time.

Figure 3: Temperature prediction using a generated physical model.

Future Developments

The AI is highly practical for detecting abnormalities in products and systems, so application to various products and systems can be expected. Toshiba will continue to expand the scope of this technology’s application to social infrastructure-related products and systems as well as verify its effectiveness, aiming for practical application by 2023.


*1: Mathematical descriptions of target events or device behaviors, based on knowledge of physics and engineering, applicable to predicting event occurrences and describing physical phenomena related to the mechanism of their occurrence.

*2: Global Information, Inc. press release, “The predictive maintenance market: An evolution to cost-effective applications”
https://www.value-press.com/pressrelease/270496

*3: A component that combines power semiconductors to integrate circuits related to power control and power supply.