Preventing the failure of large-scale facilities such as nuclear power plants! The SIAT analytics technology that analyzes vast quantities of sensor data to find anomalies
Are you all aware of how structures such as bridges or buildings and large-scale facilities like factories and power plants are managed? I've seen videos of maintenance staff on the news tapping sections one by one to inspect them, and it looks like it takes a tremendous amount of time and effort. I get worried that another spot might break before they've had a chance to check it.
While looking into this issue, I made another discovery! NEC has actually been developing a unique technology for monitoring large-scale facilities and structures such as plants by analyzing observation data from sensors. This technology is called SIAT (System Invariant Analysis Technology). From what I'm told, it can find faults and even points that are likely to fail by analyzing large amounts of data to detect anomalies! Doesn't that sound amazing?
So in this installment, I spoke with Mr. Yoshihira from the SIAT research and development team at NEC Laboratories America, as well as Mr. Asakura who is supervising research on the Japan side!
Interviewee: Kenji Yoshihira
Mr. Yoshihira is a researcher and developer for SIAT (System Invariant Analysis Technology). He researches large-scale complex systems and big data analytics technology at NEC Laboratories America, located in Princeton, United States.
Interviewee: Mr. Takayoshi Asakura
Mr. Asakura supervises research into SIAT (System Invariant Analysis Technology) on the Japan side. After joining NEC, he was responsible for the research and development of content protection and delivery. Since 2012, he has been working toward the commercialization of NEC Laboratories America's big data analytics technology.
- MitaThank you for seeing me today, Mr. Yoshihira and Mr. Asakura. Getting straight to the point, can you tell me what exactly SIAT is?
- YoshihiraBefore discussing SIAT, I'd like to talk a bit about the lead-up to its development. I think taking it step-by-step will make it easier to understand.
- MitaBy all means.
- YoshihiraAs you're no doubt aware, there are many large-scale systems(*) that operate around the clock, such as factories and power plants. Once you build these structures, there is one thing that is absolutely essential. You have to monitor it to ensure the structure doesn't fail. The collapse of a structure can have severe repercussions, so this is vital. The most straightforward way to do this is by having someone carry out monitoring, but it isn't possible for one person to watch over a site 24 hours a day, so you can't cover all your bases this way.
(* In this article we also use the term "system" to also refer to large-scale facilities and structures)
- MitaWhat do you do, then?
- YoshihiraUsually you would use sensors to detect anomalies. There are a variety of different types, such as those that pick up vibrations, and those that check pressure or temperature, and you monitor the site by looking at observation data from these sensors. Normally, anomaly detection rules are set for each sensor, with "threshold values" used to determine the presence of an anomaly when data values go above or below a certain level. Then, once these rules have been thoroughly threshed out, they are combined to detect anomalies throughout an entire system.
- MitaThat's the only way you can determine when a fault exists, I guess.
- YoshihiraHowever, the scale of systems has now grown to the point where sensors number in the thousands or tens of thousands. And the more complex a system gets, the more complicated the points to monitor and rules become. This makes it more difficult for even experts to set sensor thresholds, or in other words the rules for detection. Basically, defining what is normal is now harder, even before you consider the monitoring of facilities.
- MitaIf you don't know what the normal status is, you can't determine whether there's a fault, right?
- YoshihiraYes, and that's a problem. We developed SIAT to resolve issues like this.
- MitaI think I have a better picture of the background to its development now.
- YoshihiraWith SIAT, you begin by building a model for a system's normal state of operations. Here you can think of a model as a method of expressing a system as an abstract. If you have this model, you can compare it with the current state to detect whether an anomaly is present.
- MitaHow do you build these models? It sounds like it would require a high level of expertise.
- YoshihiraNo, all you need is measurement data from the sensors at the factor or power plant, etc.
- MitaI see! But you mentioned that factories and power plants have a huge number of sensors installed. How do you build a model from such an extensive array of data?
- YoshihiraWe feed a large volume of sensor observation data into SIAT, and build a model for the system through comprehensive automatic analysis. For example, from a large amount of data, we look for sets of sensor points with data values that fluctuate in the same way.
- MitaSensor points that fluctuate in the same way?
- YoshihiraWhen a system is operating normally, most equipment within the system works in unison, so there are actually quite a few sensors with data values that fluctuate in the same manner. SIAT finds two sensor points that behave the same way like this, and regards this pair as an "invariant" for the system under normal circumstances. This information is then stored in a computer as a relational expression.
- MitaSo it checks the relationships between sensors.