Presenso CEO Eitan Vesley explains how big data and artificial intelligence breathe new life into mechanical systems by spotting breakdowns before they happen and optimizing mechanical efficiency.
There has been much hype, much discussion, and much consternation about the role of artificial intelligence and machine learning, and are the robots taking our jobs?
For TechRepublic and ZDNet, I'm Dan Patterson, and it's a pleasure today to speak with Eitan Vesely. He is the CEO of Presenso, a machine learning and analytics firm based in Haifa, Israel. They focus on a really unique challenge. That challenge is fixing machinery, or helping optimize machinery, actual physical machines, using machine learning and artificial intelligent. Eitan, thank you very much for your time today. I wonder if you could first tell me a little bit about how you bridge that gap between software machines and hardware machines?
Vesely: Sure. Maybe I will start with a few background, or introductory words about myself. I'm a mechanical engineer by education, so most of my experience as an engineer was in industrial automation, process monitoring and control, supporting machines in production lines. And as part of my role I used to support customers with failed machines where their production line would stop. Specific machines would crash and as a support engineer, what I would do first is try to understand what happened. What is the root cause of the problem and in doing that by analyzing and looking through tons of data.
When I say analyzing, of course, it's doing it myself, manually. Sometimes using Excel. Sometimes using other types of software, but then the idea occurred together with my co-founders, why not doing it automatically the same data that far more advanced analysis, but the main difference is do it before the machine fails rather than after it fails and stops production. So, by employing or enabling, bringing machine learning capabilities or AI to the industrial market today, this is building the bridge I would say between machine learning and actual machines in the field.
Patterson: What type of machinery are we talking about? There's a broad spectrum of course, everything from automotive to industrial. What type of machines and how does your technology optimize the performance of those machines?
Vesely: There's a few distinguishable or a few categories that we can look at. We are focusing mainly on the heavy process industry in which machines are big. They don't have a spare machine in the warehouse that they can just bring in and operate instead of the failed one and there is a lot of data. Hundreds and thousands of sensors in such a process and this is our first market we're focusing on as opposed to, as you mentioned, to the discreet manufacturing, which is a bit different market segment, although it's all under industrial umbrella.
Patterson: How do you retrofit analog sensors with digital sensors?
Vesely: This is a good question. We are not retrofitting anything. We are a software solution and not a hollow based one. You're right. I mean there are solutions, which develop new sensors whether they're new vibration sensors or new acoustic sensors, and then they really have to go and retrofit new types of sensors. Looking on the other side, this is the first category of solutions. On our side, we're talking about the software base solutions. We're experts in data analytics, not in the actual mechanics or physics of the machines.
Looking inside the software solutions, they're several types of solutions and some of them are relying on low level understanding of the machine itself. Normally this is a digital twin approach. In order to monitor the machine, you really have to know how it's built and what's the physics that's happening there. On the other side of the spectrum, this is where we're at, it's a rather diagnostic approach saying, "I really don't care what your machine is doing. What is the sensor measuring? I just need the data. I need some historical data to train on and then we can predict how it's supposed to be behaving."
Patterson: So, that all makes sense, but before we move on, I don't want to stick on this point, but how do you gather information? I know it's not your core business, but how do you get reliable information from analog processing and analog feedback mechanisms?
Vesely: Yeah, so let's take one example. There is a power plant in Europe that we're working with, monitoring today. Most of the data in these process industries, most of the sensors are already connected and the information is flowing to a centralized database known as a process historian. Though there are databases storing three, five, 10 years of historical data, which is just lying there, nobody really is doing anything with it except for post mortem analysis. We're simply interfacing to that historian database and gathering thousands and tens of thousands of sensors and streaming them to our cloud. That's it.
Patterson: So, this totally makes sense and I think that there is a lot of historic data that as you referenced has be sitting around unutilized, and when you add it to machine learning systems, you can do some interesting things with it. How do you train your networks to properly tune towards one particular system versus a different client's systems? Do you use general adversarial networks? Are you using the data that you find? The historic data. How do you train systems to perform very specific task and not just general tasks?
Vesely: Sure, this is an excellent question and one of our main efforts and challenges was to develop a scalable and generalized system as much as possible with remaining agnostic to the specific asset itself. We believe that there is no one strong algorithm that can fit all cases in all sensors and with that belief, we went out and developed what's known today as a meta learning base analytic engine. What the meta learning actually means that there are a few dozen of very different algorithms used for data analysis. One algorithm could be very good at one place and be very, less good I would say, less fit in other places.
And on top of these dozens of algorithms, which by the way, they could be pure machine learning. They could be deep learning. They could be seen as processing. On top of that, what we call tool box of algorithms, we have developed the main algorithm. The mother algorithm, which is responsible for the selection of the best performing analysis algorithm. To the calibration of it and to the continuous validation of it, so all that process of selecting the tools, the algorithm you want to be working with, match making them to a specific sensor is fully automated.
Just as an example, if you take a [device] that was 500 sensors, know all the temperatures, vibration, voltages, currents, pressures, there's a very good chance that each and every one of them will be analyzed in a totally different way, which is tailor made by the system specifically to that sensor and that has proven itself. When several evaluation process our customers performed we have beat the competition. Also an internal benchmark that we have conducted and also with market benchmark where customers performed.
Patterson: How widely used are these systems or how widely used will they be in the next say 18 to 36 months?
Vesely: I can tell you that from my experience in the last couple of years or even three years in that market, the level or the rate of adaptation or adoption, is increasing and we're now at a point that if industry before was like something to speak about, now we see how the market is doing it. We have request for a query, request for information, and for quotes from people who are actively looking for this and I expect it will increase in the next 18, 36 months. It will definitely increase.
Patterson: Could leave us with a forecast for mechanical systems in the next year or so? The machines are coming to take our jobs. I wonder if we could dispel this myth. Technology has and always will replace historic jobs, but in what ways is artificial intelligence, particularly apply to mechanical systems creating jobs and creating an environment of new economies.
Vesely: Okay, so first of all, let's speak about that part of taking jobs first. Eventually, you know as a mechanical engineer every machine will need maintenance and the maintenance at least in the next decade or two decades will be done by people actually going down to the machine and replacing the part, so this is one part of the story.
With that said, of course, there's a lot of opportunities to new roles, new jobs that are being open, whether it's data scientist or application engineers that will know how to work with these analytic tools. We do see today, traditional firms that for decades are hiring mechanical engineers are now looking into developing their own software services for the machines that they sell, so begins the new, my two cents is that the need in data scientist will increase very fast in the industrial sector.
- Machine learning: The smart person's guide (TechRepublic)
- How to prepare your business to benefit from AI (TechRepublic)
- Cheat sheet: How to become a data scientist (TechRepublic)
- Executive's guide to AI in business (TechRepublic free ebook)
- How AI and machine learning can help solve IT's data management problem (TechRepublic)
- Research: Companies lack skills to implement and support AI and machine learning (Tech Pro Research)