Using AI and ML to Optimize Industrial SystemsBy Jan Hogewoning | Tue, 12/01/2020 - 11:44
Q: What does SparkCognition bring to the market?
A: The founding mission of SparkCognition is to ensure reliability, safety, efficiency and security in cyber-physical systems. We were created to solve the most difficult industrial challenges using artificial intelligence (AI) and machine learning (ML). We are making sure complex systems, such as power plants, grids, ships, planes or offshore platforms, operate safely and efficiently. We want to maximize uptime while minimizing risk.
The four products we offer are the result of that vision. SparkPredict, for example, protects companies against natural failure. This means protection against failure of a machine that is rotating, for example, and naturally degrading over time, without human interaction or digital attacks. There is also unnatural failure, where a machine is either physically attacked or attacked through a digital surface. In this case, DeepArmor protects IT systems as well as OT systems, which are the control systems that operate power plants, grids or offshore platforms.
Q: Why is artificial intelligence and machine learning crucial in these areas?
A: Challenges, such as optimizing how an offshore platform or power plant operates or going from 10,000 planes to operating millions of drones, are problems where human capabilities and expertise can be augmented with AI based solutions. These systems are increasingly complex, especially as assets age and their behaviors change, throwing more people at these problems is not a sustainable, scalable solution. You must maximize your people’s hard-earned expertise by enabling their efforts with AI based software. This is the challenge we are addressing, while improving the science and research of AI and ML, automating to such a scale that allows you to manage these complex systems.
Q: How are your models improved over time?
A: The rules-based systems that inform industrial operations today are designed by humans and humans take months or years to design these rules. The outcome is not optimal. To give you an example, when do we know the best threshold for an alarm? What happens in human rules-set systems is that you end up with too much noise, too many alarms. In a hospital emergency room, nurses know exactly which alarms to listen to. In a factory or offshore platform, however, people must react to every alarm. We are deploying systems that ensure only alarms go off to indicate important things and reduce all the unnecessary noise, including false alarms. Secondly, we want to make sure alarms are explainable. A human does not have to be an AI expect to understand and take actions in the case of an alarm.
We like to say that a human-made fixed rules-based system degrades from day one. It depends on who operates the system, who maintains the system, how the weather changes during operations, the type of fluids that run through pumps or valves and many other variables. The nature of how that asset operates alters it from day one. A fixed rules-based system cannot adapt to the new normal, whereas an AI and ML system can.
Q: Which segments of the Mexican markets are your targeting?
A: We have been in the Mexican market since January 2018. We are active mainly in the oil and gas and utility sectors, with some very large clients there. We have also been doing some work with our Darwin product to automate the model building process for detecting fraud and anomalies in financial transactions.
We are active in cybersecurity through multiple industries. That can be any industry, with DeepArmor as the main product.
Q: What is your main contribution to the oil and gas industry?
A: There are a couple of solutions that we have brought to market that are now public. One is for upstream production platforms. We are reducing downtime, caused by production impact events (PIEs). According to McKinsey, the average percent uptime of an offshore platform is somewhere in the mid-70s. With our solutions, we have increased this to the mid to high 90s. Considering that these platforms are essentially pumping cash out of the subsurface in the form of a sellable global commodity, the impact of maximizing productivity is in the tens to hundreds of millions of dollars for some of our customers.
We delivered projects where the time between getting access to the data and deploying the first model was as little as two weeks for very complex systems. There are up to 40 systems in an offshore platform. Being able to deploy with scale, in multiple systems, is fundamental in this industry. We have learned that failure is not due to one asset going down. Platforms are brought down for maintenance due to a systemic instability. This is some type of degradation at the system level. If you are deploying a model in one compressor or one system, you do not have the visibility necessary to make the necessary decisions. Our models tell the client what the contributing factors for a systemic instability are, and which sensor or what pattern leads to this failure.
A significant area of success is our ability to scale models up. Our competitors must build a model, then another one and then another one. Our solution can learn from the previous model and adapt. The solution can then be taken to other platforms.
Q: How do your solutions affect the need for maintenance in industrial operations?
A: There is less maintenance, it is optimized and different actions can be grouped together. Obviously, it is different for wind, offshore, utilities and other industrial operations. But the goal is to optimize the maintenance process.
We worked with an urban refinery in Houston called Texmark. We took the company through a digital transformation in one project. Before, when something broke at 2 a.m., for example, maintenance would get a team together and fix it. However, they ran the risk of not having the right people, with the right training and the right components to fix it. This is a safety issue, as well as an efficiency issue. Now, they have a model that can predict when something is going to fail in a few weeks. They can plan ahead, get a team together and find the right time to fix it. In addition, they can plan a number maintenance actions together in one visit, saving costs.
Q: How do you tackle cybercrime with your technology?
A: There is not enough human talent to address the problem. When only one or two types of new malware were coming out, you had the time to look at the signature and design a response. Now you are talking about zero-day malware attacks, where criminals are designing or buying custom malware on the dark web that is specifically made for your company. They attack after they have been inside your company for months. Tens of millions of malware variants are emerging every week that have never been seen before. We use artificial intelligence and machine learning to identify that never-before-seen malware. Our software has seen hundreds of millions of variants of malware and it does not need to see that specific type of malware before determining it is malware. It provides an explanation and then an expert can go in and audit it and confirm if it is malware or not. The system learns again and improves.
The old way, using human developers, has not been possible for the last five years. This is why you see very mature operations, large institutions, that are unable to protect themselves. We have to ask ourselves: how can a bank lose all its customer data? How does a credit bureau expose hundreds of millions of records on the web? How does the US government lose all the data of its employees? If it was addressable by current methods it would have been addressed. Here in Mexico, last year, an oil and gas company was attacked in this way.
Q: What is behind your collaboration with Siemens in industry cybersecurity?
A: Cyber-physical systems can suffer unnatural failure by someone injecting malware that alters the operations of the system or extracts data or intellectual property. Normally, the goal is to extract money from the company, known as ransomware. Siemens, like us, believed that this is a problem that cannot be addressed on a human scale. The company decided to partner with us to develop better solutions for operational technology systems. OT systems have connected and disconnected systems. You have legacy systems, which have been operating in industries for a long time. If you try operating modern anti-malware software on these systems, it is not going to work. We deployed a system built for the reality of today. It is trained with hundreds of millions of malwares and can operate in an architecture that contains new and older technology that has both connected and disconnected components.
Q: What sectors represent opportunity for SparkCognition in Mexico?
A: Early adopters who want to use AI to drive business transformation and accelerate growth in the years to come as industries, even those built on physical assets and machinery who view their businesses through a digital first lense. Where decisions are informed by data and enabled by AI models. Initially, making recommendations to human operators but gradually automating repetitive tasks and providing insights for specialized actions. Implementing AI can be a comparative advantage right now. In the next five to 10 years, it is going to be the standard. The benefits are hyper accelerated. Often, what is required is open-minded leaders.
SparkCognition uses AI and machine learning models to provide reliability, safety, efficiency and security in cyber-physical systems. Its primary sectors are oil and gas, energy and aerospace.