Toshiba announces AI technology that can detect abnormalities in equipment and determine the cause
Toshiba announced that it will develop artificial intelligence (AI) technology that can detect abnormalities in equipment and determine the cause of abnormalities, aiming for practical application in 2023.
The company automatically generates a highly interpretable physical model (digital twin) that can express the state and operation of the device from the time series data of the device, and in addition to the judgment of "unusual", the judgment of "why is different". It is said that it can be used for predictive maintenance of infrastructure equipment because it will be possible.
Specifically, by utilizing AI, which is an original algorithm, in addition to the conventional mechanism of detecting the temperature rise of the device by using sensors etc., "The temperature rise is caused by clogging of dust. In addition to facilitating improvement and planning of countermeasures, it is possible to utilize the mechanism of abnormality occurrence for the maintenance of complex infrastructure equipment.
This technology was developed by the Intelligent Systems Laboratory of the Research and Development Center, and is highly practical in detecting anomalies in products and systems, and is expected to be applied to various products and systems. In the future, we will expand the scope of application to social infrastructure-related products and systems, and proceed with verification of their effectiveness. "As an infrastructure service company, we can expect to improve the reliability of equipment and contribute to the resilience of social infrastructure by putting this technology to practical use."
Technical featuresThe physical model automatically generated by this technology expresses the correlation between data items in a network, and combines the relationships between items with functions based on physics and engineering to form simultaneous equations. Toshiba's original AI algorithm is used for the combination of functions. Until now, designers with specialized knowledge and know-how will use AI to automate the optimal physical model that was constructed by combining functions like a puzzle.
According to the company, the image is to think of a network connecting points to which physical quantities are assigned, and to automatically determine which point is related and what kind of relationship it is with its own AI. It will be a point of development. If you replace it with a region, you can understand the relationship from the point difference, such as the temperature of Tokyo and Kanagawa is related, but the temperature of Tokyo and New York is not related. To derive.
Function candidates utilize the function database accumulated based on the knowledge of mechanical engineering that Toshiba has cultivated over many years. In addition to temperature changes, it is said that horizontal expansion is possible by utilizing functions for changes such as abnormal noise, friction, and wear, and it is a major feature that a physical model can be generated quickly and easily.
The feature here is that we have developed and used three unique AI algorithms. The three AI algorithms are categorized into AI that selects the correct combination from a huge number of function candidates and AI that performs efficient coefficient estimation, all of which are technologies aimed at analyzing big data such as neural networks and deep learning. Instead, the approach is to use a small amount of data to create a highly accurate model.
Technical featuresIn the conventional technique, it was difficult to efficiently combine a huge number of functions without changing the physical meaning of the function, but a new sparse estimation algorithm that can correctly consider the degree of physical influence of the function is utilized. bottom. Furthermore, by combining a spatial search algorithm that efficiently selects a huge number of function candidates, and a data expansion algorithm that enables highly accurate prediction by adding preprocessing that differentially integrates sequence data in consideration of time constants. , Developed a new AI technology. As a result, it was possible to automatically generate a physical model with excellent interpretability.
Furthermore, there is no need for data on the dimensions of equipment and physical properties of parts, which was required to generate a conventional physical model, and there is also the feature that a physical model can be generated using only measurement data from sensors. This makes it possible to periodically update the physical model during the operation of the product or system, and by analyzing the changes in the updated physical model, it is possible to detect signs of abnormalities in the product or system and their causes. It became possible to identify.
The company has shown a concrete example of anomaly detection with a power module. A power module is a component that integrates circuits related to power control and power supply by combining power semiconductors. Temperature prediction is important for anomaly detection.
With the new technology, it is possible to automatically generate a physical model applied to the power module, and it was confirmed that the heat transfer form in which heat is transferred from the heat generating chip to the cooler and the heat is dissipated from the cooler by the air cooling fan is correctly selected. And that. It is assumed that the generated physical model has an average error of less than 1 degree and can predict the temperature change with high accuracy.
In the past, detailed numerical simulations that took thousands to tens of thousands of times longer were required, but with this technology, calculations that take one to two days can be completed in seconds. It will be possible to realize real-time predictive maintenance.
Technical featuresIn order to continue to use products and systems safely and securely, the importance of predictive maintenance that detects deterioration and malfunction of equipment in advance and manages them in the optimum state is increasing. According to the company, the predictive maintenance market is in the growth stage, growing rapidly from about $ 6.9 billion in 2021 to a compound annual growth rate of 31%, reaching about $ 28.2 billion worldwide in 2026. ing.
According to the company, "Many infrastructure devices have complicated mechanisms for generating abnormalities, and it is difficult to formulate improvement measures using only anomaly detection technology that presents differences from the conventional normal state. Supporting social infrastructure. In addition to reducing interruptions and downtime due to failures, equipment can minimize maintenance costs. We will propose it as a highly accurate predictive maintenance technology. "