ML21225A709
ML21225A709 | |
Person / Time | |
---|---|
Issue date: | 03/11/2021 |
From: | Office of Nuclear Reactor Regulation |
To: | |
References | |
NRC-1420 | |
Download: ML21225A709 (29) | |
Text
Official Transcript of Proceedings NUCLEAR REGULATORY COMMISSION
Title:
33rd Regulatory Information Conference Technical Session - TH22 Docket Number:
(n/a)
Location:
teleconference Date:
Thursday, March 11, 2021 Work Order No.:
NRC-1420 Pages 1-63 NEAL R. GROSS AND CO., INC.
Court Reporters and Transcribers 1323 Rhode Island Avenue, N.W.
Washington, D.C. 20005 (202) 234-4433
1 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 UNITED STATES OF AMERICA NUCLEAR REGULATORY COMMISSION
+ + + + +
33RD REGULATORY INFORMATION CONFERENCE (RIC)
+ + + + +
TECHNICAL SESSION - TH22 ANALYTICS, MACHINE LEARNING, AND ARTIFICIAL INTELLIGENCE FOR NUCLEAR POWER PLANT ACTIVITIES
+ + + + +
- THURSDAY, MARCH 11, 2021
+ + + + +
The RIC session convened via Video Teleconference, at 10:30 a.m. EST, Jeff Baran, NRC Commissioner, presiding.
PRESENT:
JEFF BARAN, NRC Commissioner JEREMY RENSHAW, Senior Program Manager, Artificial Intelligence, Electric Power Research Institute MATTHEW RETZER, Senior Manager Equipment Reliability, Exelon Nuclear MARIA LACAL, Executive Vice President & Chief Nuclear Officer, Arizona Public Service Company, Palo Verde
2 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 Generating Station KIMBERLY WEBBER, Director, Division of Systems Analysis, RES/NRC
3 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433
4 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 P R O C E E D I N G S 10:31 a.m.
COMMISSIONER BARAN:
I will be moderating today's session.
We'll be talking about how advances in computing technologies and access to both historical data sets and real-time data collection may lead to the expanded use of artificial intelligence, or AI, and data analytics in nuclear power-plants.
We're going to hear about what's happening now in this area and what the future might look like. And we'll be discussing the potential safety improvements that could result in predictive maintenance in other aspects of plant performance.
To lay the foundation for our discussion, our four expert panelists will each give a five-minute opening presentation. Because Q&A is the fun part, I'm going to strictly enforce the five-minute limit.
That will leave us with about an hour to discuss the issues you and audience are most interested in. As you think of questions, please submit them through the RIC platform.
To get things started we have some pre-
5 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 prepared questions, but we're counting on the audience to come up with the questions to sustain a great discussion. We'll work hard to get through as many questions and topics as we can.
For most questions I'll give each panelist an opportunity to share his or her thoughts.
Let me start by introducing our terrific panel. Dr.
Jeremy Renshaw is the Senior Program Manager for Artificial Intelligence at the Electric Power Research Institute.
He leads the AI.EPRI initiative focused on AI and machine learning technologies and their potential applications in the electric power sector.
Matt Retzer is the Senior Manager of Equipment Reliability at Exelon, where he focuses on the maintenance rule, surveillance frequency control, and risk management.
Maria Lacal is the Executive Vice President and Chief Nuclear Officer for the Arizona Public Service Company. She'll share how analytics, AI, and machine learning are being used at the Palo Verde Generating Station.
Rounding out the panel is Dr. Kim Webber, who serves as the Director of the Division of Systems
6 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 Analysis in NRC's Office of Nuclear Regulatory Research.
Jeremy, let's get started with your presentation?
MR. RENSHAW: All right, thank you very much, Commissioner Baran. I'm very excited to be here with you today to talk a little bit about some of the efforts that we're doing here at EPRI related to artificial intelligence.
So if you want to go to the next slide?
All right, very good. So here at EPRI we really are looking at three major areas that we're focusing on for artificial intelligence.
The first is really around building an AI and electric power community, and so some of the things that we're doing here you can see on the slide.
I'm not going to go through and read these, but the main items that I want to focus on are, really, that we're looking at bringing the two industries together to understand the needs and limitations of each other.
And you can think about this a little bit in terms of if you have an AI algorithm that's running on a website or targeting some advertising, and it's
7 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 only 90 percent effective, that's generally pretty good. If you have an AI algorithm that's running a nuclear reactor and you only have a meltdown 10 percent of the time, that's incredibly bad. So we need to help understand the difference in reliability and safety that we have in the nuclear industry compared to what the AI industry is generally accustomed to.
So some of the things that we're doing to be able to build this community are hosting a number of events, the first one will actually be next week on March 18th, where we're bringing together a very exciting executive panel from across these two industries to talk over what some of the needs and issues are.
The second area is really around data collection,
- curation, sharing, and solution development where we're bringing in a wide range of data sets at EPRI to be able to cleanse and share those with the AI community to be able to build various AI and machine learning techniques and solutions for some of the big data challenges that we have in the industry today as well as facilitating future demonstrations.
8 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 The next area is really around deepening AI experience within the industry and not just nuclear, but the entire electric power industry.
So some of the things we're doing here are publishing white papers and landscape reports, developing training through our EPRI-U, integrate with data initiatives, as well as engaging stakeholders across these two areas to help them understand and interact with each other through conferences, meetings, and workshops and other areas.
So if we can go to the next slide, there are really four areas that we see that need to be successful for AI to be implemented in the power industry. And the first one is bringing together the data.
So data is really the fuel that is the fire for AI, so we need to have large, clean data sets that we can use. The second area is having a power industry that understands AI. How does it operate? What are some of the basics of AI? And really, how to engage the AI community effectively.
The third area is really an AI community that understands the power sector and our needs and the limitations that we have as well as an
9 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 understanding of the physics and the data.
And
- finally, understanding AI performance in terms of how do we apply it to be effective? What are some of the internal biases that we could come across, as well as having explainable AI.
And so if we go to the next slide or if we hit forward, you can see we're looking at EPRI, trying to be kind of in the middle of this and bringing these industries and these areas together.
All right, so if you hit next again, we're looking at essentially a timeline of some of the events that we have planned. I won't go into these in any great detail, but you can see we have a number of events to help build this community and bring people together.
So we have several of these over the next several months and we would invite people if they're interested to get involved in these. The March event is next week, like I said. The registration is open now, and you can go to AI.EPRI.com to see that.
So if we go to the next slide? These are some of the data sets we're looking to bring together, and these are what we're calling the EPRI 10 or the
10 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 ten most data-intensive challenges that we have in the entire electric power industry.
So I won't go into any significant detail on these either. You can read these; they're in the slide materials that are provided to you. If we go to the next slide, you can see how these fit into a number of different projects that we're hosting at EPRI.
So we have about 50 different projects in AI at EPRI, 20 of them are nuclear specific. We have about 30 that I'm showing here that are funded by our AI initiative. They were really seed funded to get a lot of different work going across, again, like I said, all of the electric power industry, not just nuclear.
And then to round it out, if you go to the final slide, these are a number of reports and white papers that we put out. All of these are freely available, and if you go to our website AI.EPRI.com you can see about 40 different reports that we put out that are available to different members of the public, so you can go there and read more about these.
So I look forward to taking a number of questions later on and talking more about what we're
11 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 doing with AI, as well as bringing together utilities and the AI community. So thanks, Commissioner Baran, and I'll pass it back to you.
COMMISSIONER BARAN: Great, thanks, Jeremy. Matt, you're up next.
MR. RETZER: Great, so I am Matthew
- Retzer, I'm the Senior Manager of Equipment Reliability from Exelon Nuclear. And I want to talk about three different ways we are using analytics and AI in our equipment reliability programs.
So next slide? The first area is around maintenance rule functional
- failures, the terminations and issue report reviews. Our need for the analytic really stems from Exelon has 22 units, we run 12
- sites, each of which generates approximately 100 to 200 issue reports a day. So this is a large data set, and what we're interested in is finding maintenance rule functional failures.
So these are events where our safety-related or important equipment failed to operate as we expect. And really, that only -- that is a very small population, approximately 0.1 percent of all those issue reports had that data in it.
So, traditionally, about once a month, an
12 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 engineer would sit down at their desk and they would review a large pile of these (audio interference) electronically and have to answer questions, pretty much an administrative burden.
So looking at that data set, you know (audio interference) most of the data comes from (audio interference) form. There's fields that (audio interference) fields and (audio interference) fields (audio interference) like a block field, but typically, in nuclear power we find -- use a lot of consistent language.
So we felt there was a very high probability of success that we could create an analytic that would get us to take that large population information and boil it down to things that we really are concerned about and then, focusing on those, find our actual maintenance rule functional failures.
So next slide please? So the first thing we did was -- and I won't go too much into detail about the AI and analytics because I know there's other people in this panel who are experts there.
But we had to train the model where we were able to determine which IRs or which issue
13 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 reports would be a maintenance rule functional failure.
So we're fortunate in that we have several hundred thousand of these issues in our database that we could reference, and we also have a very accurate database of exactly which one of those resulted in failures.
So, basically, we used like a process of clustering to go through these IRs, review them, and determine which ones are actually maintenance rule functional failures.
Part of that was using natural language processing, so we would look at this blob text information and section it out into simple statements. I give the example here of Jack went to the store and bought food.
So with natural language processing you can split that up to a person shopping for food and then fed that into an artificial neural network to eventually get us the outcome that we were looking for. And then we have to train that model over thousands and thousands of iterations.
Next slide, please. So this is basically the flow of information and how we use our maintenance
14 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 rule analyzer. We bring in inputs, the text field that I was talking about for each one of our issue reports with all of these scripted information in it, and then we also have binary fields, which are like the yes-no fields I was discussing.
That information comes in, we pre-process it, clean it up, and then we use this Bayesian text confidence, which is just a complicated way of saying we look at the word combinations to get an idea if those word combinations could reflect a potential failure or not.
We feed that into -- and we get a score, and we feed that into our neural network. And then it basically classifies things as either being a maintenance rule functional failure or not and gives us a confidence level for each one of those ratings.
And then based upon that rating and the confidence level, we will then forward that to the engineers to screen in a separate program.
So, effectively, we take hundreds of thousands of issues reports over the course of a year, and we're able to concentrate it down to less than 20 percent. And we're running about 10 percent of those right now still require manual reviews.
15 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 Next slide, please. The second analytic or machine learning process we use at Exelon is for our maintenance and diagnostic center, that is our monitoring center. That is a corporate function where we bring in a lot of field data and basically try to monitor our different 22 units.
I gave a little bit of an example here for this slide. You'll see a simplified feed pump turbine. You can see some of that information from the field goes to our plant process computer, and that'll go directly to the control room.
In some cases, we can't do that between cybersecurity rules and just equipment limitations.
So we take some of that data and we might just send it to a server, but we take all that data and eventually get it to our M&D (phonetic) center.
And then we associate that data with different tags to a specific component or a template, and then with that comes a model where we get expected behaviors. So when some temperatures go up we expect maybe pressures to go down. And if we see deviations from that behavior, we go in and get alarms and warnings, as you can see there.
And then we, of course, this model we can
16 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 train so if we're getting false alarms and that type of stuff, we can go back and we can train the model and make it better.
And if we go to the next slide, the last thing I wanted to highlight was what we used for maintenance strategy optimization. And this is where we look at our maintenance activities, and we effectively score it.
As I show in this diagram, we use our PM-to-CM cost ratios on one axis and then aggregate score on the other axis, where that aggregate score is based upon our as-found conditions, the frequency of our work, the amount of that corrective maintenance we have to do versus preventative, the criticality of the component.
And we also use that natural language pattern recognition to go ahead and look at the as-left comments that the maintenance personnel used.
And that all rolls up into this aggregate score.
And in this particular plot we focus on the very top right section and the very bottom left section. In the top right section, these are things that we probably do too much maintenance on for preventative maintenance, and you see I highlighted
17 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 in this slide that we have circle sizes, basically, that show us the costs.
So that gives us some items that might be an expensive cost item that we might be doing too frequently that we can do less frequently, which also has the added bonus of preventing maintenance-induced failures or reducing that.
And then the bottom left of this plot would be things that we're doing too much corrective maintenance for and not enough preventative maintenance. And so those are items that we need to improve our preventative maintenance strategies for.
So that's our three different tools that we use at Exelon, and I'll turn it back over to you, Commissioner.
COMMISSIONER BARAN:
- Thanks, Matt.
Maria, you're next.
MS. LACAL: Thank you, Commissioner Baran, and much like Matthew, I'd like to share a few examples where our team is using artificial intelligence, machine learning, and data analytics.
They're really helping us to be much more efficient in the way we do business, and it really allows us to focus our valuable resources on the more
18 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 safety-significant issues.
So if you go to the next slide, please?
On this slide you can see a number of applications that we're using. So, for example, our equipment anomaly detection system, that was developed in house by one of our engineers with knowledge of the plant and a real passion for data analytics.
We teamed him up with one of our IT folks that also has a passion for artificial intelligence and machine learning. So the result of this is the creation of a system that handles plant data at a nuclear scale using data analytics.
So this system detects abnormal plant behavior and allows us to focus on potential equipment malfunctions prior to failure at the system level scale.
So abnormal combinations of patterns over hundreds or really thousands of process point interactions such as temperature, pressures, flows, similar to what Matthew talked about.
They're displayed in a really user-friendly manner for further investigation, so the result is a lot of efficiency gains as well as proactively addressing abnormalities before they
19 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 become plant issues.
Our Data Ingestion and Network Analysis tool are what we refer to as DIANA. Good NUCs (phonetic), no lack for acronyms. DIANA accelerates an analyst's ability to put pieces together after cognitive training during an event, or perhaps detecting connections in the data that may not have otherwise been uncovered that utilize graph analysis and machine learning.
So, for example, the use of this tool by planners eliminates the need to dig through tons of old work orders to get a list of applicable parts.
Instead, DIANA readily recommends the parts for a given work order using the artificial intelligence recommendation engine.
Kind of the same way that Netflix recommends movies to all of us, DIANA can recommend parts to work orders and planners. From work management automation and optimization perspective, advances in process automation allows us to efficiently balance the workload on our craft.
Each and every work week our program runs hundreds of simulations of how long our work takes in a work week, and it can detect if we're over or
20 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 underloaded in a given work group.
So schedulers can see at a glance how overloaded or underloaded the schedule weeks are, as compared to historical standards. This provides us with better loading capabilities. It allows us to better use our labor resources, so planners no longer need to like manually pull items from a sand box and place them in the right time or day and kind of work through that to schedule the work week.
This artificial intelligence uses techniques to match the correct work item into its time, day, and week slot in the repeating backbone schedule.
And finally, the use of machine learning and artificial intelligence in our corrective action program really allows us to automatically screen the condition reports, auto-generate corrective actions, code the corrective actions using artificial intelligence.
And so where artificial intelligence is taught how to understand English by Google, for example, at Palo Verde we're teaching it how to more deeply understand nuclear dialect, kind of a natural language processing that Matthew talked about. But
21 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 our data science team calls it the nuclear natural language processing.
Artificial intelligence can automate repetitive, low-impact knowledge work and really lets our condition reporting screening committee members focus on the more complex items that really need their attention and their experience.
So in recognition of all of these innovative projects that our team has implemented, I'm really proud to say that we've received the NEI's Technology Innovative Practice Best of the Best Award for process automation using machine learning.
So thanks for the opportunity, and, Commissioner Baran, back to you.
COMMISSIONER BARAN: Thanks so much, Maria. Kim, you're next, thanks.
MS. WEBBER: All right, great, thank you very much, Commissioner. It's my pleasure to be a part of this esteemed panel. These are really exciting times at the NRC as we transform into a modern risk-informed regulator.
I'll quickly describe how the NRC is exploring and employing advanced technologies, like data science theories, artificial intelligence, and
22 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 process automation as a
key part of our transformation.
Next slide, please. The thought of applying artificial intelligence to nuclear power plants and applications dates back to at least the 1980s.
With more recent advancements in computing platforms and algorithms, combined with access to larger amounts of data, the potential for using artificial intelligence and process automation really began to emerge, especially in the context that Matthew and Maria described earlier.
It's important to the NRC to understand the various technologies and how they can be applied and used in the nuclear industry and to ensure that they are done so in a manner that supports public health and safety.
Additionally, we are looking at ways to use these technologies within the NRC to improve our own efficiency of our regulatory oversight responsibilities and to make our information available in more meaningful ways to our stakeholders.
Next slide. To get better acquainted
23 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 with advancements in this area, we talked to other organizations including a few electric utilities, the Electric Power Research Institute, the Department of Energy, and other federal partners such as the General Services Administration, to learn about their use cases and their applications of these technologies.
Next slide? Thus, over the last year or two, the NRC has been learning about and has begun using data science theories and practices such as data warehouses, robotic process automation, and artificial intelligence including machine learning and natural language processing.
Next slide, please. A key part of our journey is building organizational capacity through staff training and, equally important, through the development and implementation of use cases.
For example, over the last several months, we have been evaluating the capabilities of IBM Watson and other vendors' products to perform ADAMS searches to retrieve documents with a high degree of relevance, much like Google's search capability.
The application of natural language
24 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 processing to unstructured data housed within the ADAMS repository could provide enormous benefits to both internal and external stakeholders.
Using robotic process automation techniques, we developed a bot to automate entering contract invoice information into the NRC's financial systems, thus eliminating the need for manual, repetitive, human data entry by the NRC staff.
And also, we are learning how to apply a technology platform called no-code/low-code, a
business process automation approach to facilitate the intake, review, and disposition of work requests within the Office of Research.
Next slide. As our understanding of and experience with data science theories, artificial intelligence, and process automation grows, possible applications are shown on this slide and include the use of data analytics to make NRC's data easily accessible to decision-makers and to the public.
This afternoon there's a RIC session that will describe these activities in more detail. We're also investigating the development of what we call RESBots to use as research assistants to aid NRC researchers in data-mining information contained in
25 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 a wide variety of the NRC's documents and databases to conduct our research more efficiently.
We're also investigating the use of autonomous vehicles such as drones that could aid in site decontamination activities. And we'll also be investigating the use of artificial intelligence techniques to simulate nuclear power reactor operating cycles.
Next slide. Lastly, the NRC must be prepared to make informed decisions on the nuclear industry's safe use of these technologies. At the NRC, we are only limited by how far our imagination can take us when it comes to using these technologies internal to the NRC.
We're striving to transform our processes and enable our staff to be better risk-informed regulators by increasing their access to all of the NRC's data and information.
Thank you for your attention, and I look forward to taking your questions. Back to you, Commissioner.
COMMISSIONER BARAN: Thanks so much, Kim. All right, so thank you all for your opening presentations. We now have a lot of time for
26 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 questions.
The questions are already starting to come in from the audience, which is terrific. Keep them coming. Let me just start with a couple of big-picture questions.
And maybe for this first question we'll go in the order of the presenters, the presentations and then after that maybe kind of cycle it around?
How do you see artificial intelligence and data analytics providing a positive safety benefit for nuclear power plants?
Do you want to start, Jeremy?
MR. RENSHAW: So I would say that there are a large number of ways that AI and ML technologies can improve safety at nuclear power plants.
And I would say the two biggest areas that come to mind, first off, automating time-consuming tasks which allow plant staff to focus on more the safety-significant items in terms of -- I'm sure anyone listening today can think of items that are taking up time in their day that could be automated or done more efficiently.
Just in my past week I remember I spent about an hour looking for a specific document within
27 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 our system, and I thought boy, wouldn't it be nice if I could just ask Siri or Alexa or something to just go find this for me and I wouldn't have to spend the time?
So automating time-consuming tasks could help provide a lot of value right away to us. And second is really multiplying the effectiveness of
- staff, essentially giving them superpowers by allowing people to do what people do best and machines to do what machines do best.
So if we think of people, we're very good at complex analysis, developing strategies, or performing the actual work. Whereas, machines are very good at things like data processing, identifying patterns and anomalies, essentially number crunching.
So as we separate out and allow people to do what people do best and machines to do what machines do best, it really increases the safety and allows us to focus on those items that are most safety-significant.
COMMISSIONER BARAN: Matt, do you have thoughts?
MR. RETZER: I think Jeremy said it very well. The two things I was going to add to him, just
28 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 (202) 234-4433 as I discussed in my presentation, was we can use our analytics both to save our companies money for when we look at preventative maintenance items, but we're also using it to make sure that if we are having repetitive failures or concerns, we can flag that early to go address that with our maintenance strategies.
And the other thing about positive safety, we do use machine learning for our training of our plants' performances so we can flag things early. And we train the model so that we're in off-normal conditions.