The strengths of Toshiba Digital Solutions' "SATLYS" analytics artificial intelligence (AI) do not lie in its innovative deep learning platform technologies alone. One of its primary allures is that it combines our exceptional analysis technologies, refined through our use of raw, real-world data, with our AI usage technologies, in the form of a professional service that encompasses everything from data preparation to the feedback of analysis results to worksites. Let's look at our unique analysis methods, used by our analysis models, matched to our customers' operation issues, to verify the problem resolution results of inferences made by AI. This important analysis method provides support for the SATLYS service, which leverages our extensive knowledge and has been verified in applications such as abnormality detection in thermal power plants.
Through advances in machine learning, such as deep learning, AI is said to have been making strides towards application in industrial fields. However, can it be truly useful? What kinds of AI will be needed to link the digital transformation brought about by AI to customer business transformation?
In order to address concerns such as these, Toshiba's "SATLYS" analytics AI, based on platform technologies fully optimized for industrial applications, was launched in October 2017. Our goal is for SATLYS to become the leading industrial AI, and together with Toshiba's "RECAIUS" communication AI, to become synonymous with AI.
We have prepared professional services, starting with digital consulting, for our customers interested in deploying analytics AI (Fig.1).
At the core of these services is our unique SATLYS analysis method (Fig. 2).
AI is capable of processing massive amounts of data, on a scale impossible for people, to more efficiently discover new value for businesses. It has the potential to produce findings which would be difficult even for veteran experts and craftsmen, promote further creation, and continually transform business models.
However, selecting neural networks and tools and applying them to in-house systems is not enough to effectively implement measures using AI. It is of primary importance to optimize the AI to individual customer businesses and AI deployment objectives from the deployment planning stage, such as by identifying the data that will be required to perform analyses and the methods used to gather that data. AI decision accuracy requirements – whether results need to be almost 100% accurate, or if 70% accuracy is sufficient – also vary depending on operation issues and applications, the data to be analyzed, and overall scenarios. Unless scenarios are sufficiently matched to operation issues, required inference capabilities are defined, and those capabilities are sufficiently verified, even if AI is deployed it will not produce the effects desired, preventing companies from creating value linked to business transformation.
In providing SATLYS, we thought about how we could provide the extensive know-how we have accrued through our numerous AI verification tests to customers in a versatile form.
We standardized our experience in planning and implementing a wide range of AI usage, working with raw data and feedback from actual worksites. We collected our issue organization, hypothesis creation, analysis logic formulation, analysis model design, and analysis model result measurement into a single, comprehensive analysis method. This method could be used from the digital consulting and AI deployment preparation stage to create detailed AI scenarios that would be of true value to customers' businesses, carried out verification, and optimized AI inference capabilities. Our goal was fully prepared, full-fledged deployment.
This SATLYS analysis method is composed of two phases. During the first phase, the essential nature of customer issues is grasped by organizing their overall structure, and issues are identified. During the second phase, AI produces resolution measures and their effects are verified.
During the issue identification phase, the analysis method fully leverages the diverse knowledge we have developed in a wide range of worksites, such as societal infrastructure and semiconductor worksites. Customer issues and the data that will be required to perform analyses are identified, analysis methodologies are selected, scenarios are created using AI, and targets are set.
These scenarios are then used to design an optimized analysis model. Actual analysis and effect verification is performed during the following phase. This phase uses our unique deep learning technologies and our expert knowledge in broad-ranging system design, operation, and maintenance. Inference capabilities are verified in-depth, bottlenecks are eliminated, and preparations are made for actual deployment.
The superb, innovative qualities of the SATLYS analysis method, which brings together Toshiba's expertise, are already starting to be applied to societal infrastructure fields. One of these is high-precision detection of abnormalities in thermal power plants.
Thermal power plants must constantly maintain safe and efficient operation environments that provide people with a stable supply of power. They cannot suffer power supply interruptions due to unplanned power outages. The prevention of problems such as plant failures by detecting abnormalities is even more important for thermal power plants than in other industrial fields.
Plant monitoring is performed using temperature data, pressure data, flow rate data, and the like gathered from sensors. It is impossible for personal experience and gut instinct alone to accurately identify the signs of imminent abnormalities in massive systems composed of hundreds of thousands of components and numerous physical aspects, such as chemical, thermal, mechanical, and electrical elements. As a result, traditionally components are replaced at fixed intervals, regardless of whether they are exhibiting symptoms of imminent failure. Because of this, maintenance costs are a major issue.
To solve this problem, we used the SATLYS analysis method for abnormality detection in a thermal power plant to try to detect abnormalities with a high level of accuracy.
Generally, the physical model approach, in which a physical model is used to infer the states of equipment and components, requires the creation of a detailed dynamic model of the complex physical phenomena that occur within a plant. It also involves a massive volume of computation to make inferences. Detecting abnormalities using a data-driven approach, on the other hand, requires a tremendous amount of data, and the structure of the resulting model is difficult to understand, making it a so-called "AI black box."
This is why the approach we selected was the creation of a hybrid model that fuses the physical model and data-driven approaches (Fig. 3).
We built a thermal power plant "digital twin" dynamic AI model in virtual (digital) space that uses 270,000 items of data acquired from roughly 3,000 sensors to pinpoint transitions in internal conditions and changes in sensor values immediately preceding equipment stoppage in the form of feature quantities. We defined thermal power plant operating conditions, such as water supply quantities and fuel, and power generation output targets, and analyzed the model's behavior. As a result, we determined that it may be possible to detect pipe leaks in the plant's main steam and water circulation systems before they could be detected on-site by maintenance and inspection personnel. There is also potential for this model and actual data to be used in considering optimal operating conditions, such as detecting a variety of abnormalities and reducing fuel consumption.
Take advantage of Toshiba's "SATLYS" analytics AI, which combines AI platform technologies and analysis methods developed in the field, as a reliable professional service that solves the various problems you face.
* The corporate names, organization names, job titles and other names and titles appearing in this article are those as of January 2018.