Algorithmic machine learning enables predictive maintenance and optimization
April 26, 2018 | by Dr P Bangert
Outdated systems have stored a large amount of process data in their files. Machine learning can transform this data graveyard, normally used for little more than drawing diagrams, into a dynamic mathematical model of plant behavior. This model is used to calculate when equipment will fail in the future and how to change setpoints to improve plant performance. Algorithmica was founded thirteen years ago by mathematician Dr. Patrick Bangert, who felt a strong desire to bridge the gap between the world of academic mathematics and the world of industry. Its tagline "simply smart" reflects the desire to shield its customers from the complexities of math and provide simple answers to very complex questions using smart methods.
This is the April 2018 cover story onDealing with aging plants.
The immense potential of mathematics, especially machine learning, is what drove Dr. Bangert to resign as professor of mathematics at the Jacobs University of Bremen (Germany) and consider how he might actually use applied mathematics outside of the university context. "It's hard to get it out of the ivory tower and into industrial practice," he says. The benefits of implementing machine learning in chemical, power generation, and oil and gas plants are enormous. “The first thing that really came to the fore was the predictive maintenance use case. Basically this means that with machine learning you know exactly when an asset fails.”
K+S case study
K+S sees itself as an independent, customer-oriented minerals company for the agricultural, industrial, consumer and municipal segments. It operates numerous plants for the production of potash and salt products. Potassium mineral cannot be used in its raw form and must be refined. As a final step in the refining process, potassium chloride is washed to achieve the desired quality of the final product. It is then dried and stored for shipment. It is precisely this washing process that is optimized by the algorithmic software. Depending on the quality of the raw material, K+S has to adjust various process values. Also, part of the process is cyclical. Until now, operators have relied on their expertise to control the process. The software does an even better job so that the final product's quality criteria are always met and, at the same time, production is optimally cost-effective. Mr. Carsten Laukner is a program manager in the "Digital Transformation" department at K+S AG in Kassel, Germany. He says: “Using algorithmic process optimization technology, we were able to see significant improvements in process stability and product quality at our potash plant in Unterbreizbach. It was provided easily and quickly and was accepted by the operators as a useful guide.”
Condition monitoring considers each measured value individually and decides whether it is too small or too large compared to a specified threshold value. This is known to lead to many false alarms as well as conditions where the system fails but does not raise the alarm. Furthermore, the effort involved in setting limit values is laborious and bureaucratic.
Machine learning provides holistic analytics that quickly and accurately reports the current state of the machine and predicts until "next Thursday at noon there's an outage." By detecting a failure before it occurs, both the plant's availability and its total annual production can be increased, while the maintenance budget can be reduced.
With many plants operating beyond originally planned capacity, maintenance is a real challenge. "I would say maintenance is the biggest challenge for old assets," says Dr. Bangert. “The older the system gets, the more often systems fail. And then you're faced with the problem of having to fix it.” Algorithmic tools can tell in advance if an unhealthy condition or failure is imminent. “For process plants, this means they can go without preventive maintenance. They can also get rid of all the fire suppression methods that have to be put in place waiting for a failure to occur. But above all, and this is the main advantage for the customer", emphasizes Dr. Bangert, "collateral damage is avoided. If you can fix an asset before it breaks, it will cost X. If you wait until it breaks and then you fix it, it will cost 10 times as much”.
This methodology has been verified in many process industry plants on a variety of rotating and non-rotating equipment such as gas, steam and wind turbines, compressors, pumps, heat exchangers, distillation columns and valves. It can be tested quickly and easily in any plant, as it only requires installing software and creating models with historical data.
Machine learning models are also used to improve plant performance. “A process plant has setpoints at which the human operator controls that plant, and it is the operator's job to change the setpoints as they see fit to respond to influences from the outside world,” says Dr. Bangert. “The outside world usually offers two main influences: one is the weather and the other is the raw materials that go into the factory. They change in quality and composition and also vary from supplier to supplier, so factory behavior must be adjusted to take this into account. The human operator does this based on his experience and does a very good job, but not the ideal job.”
What you also see in a factory is the shifts change every eight hours. “The facilities are open 24/7, so you can have up to eight different groups of people responsible for your facilities,” continues Dr. Hit Away. “And every time the shift changes, the configuration changes too, but the installation is large, so it takes a few hours for those changes to translate into a stable situation. Eight hours later, however, the next shift arrives and changes everything again. So the system is constantly changing and never works optimally. What you need is a consistent operating philosophy that can operate 24 hours a day, and for that you need a computer program.”
Mr. Puethe, Engineering Manager at Ashland, explains: “Ashland uses the APO optimization suite at its formaldehyde production plant in Marl, Germany to maximize the bottom line and avoid foaming. By changing the benchmarks for various elements of the process, there is great potential to increase profits.”
smart soft sensor
In addition to predictive maintenance and optimization, algorithmic offers a selection of other solutions. One of them is the "intelligent soft sensor". While this may seem like a tangible thing, it is, again, a formula. dr. Bangert explains: “Sometimes it is difficult or very expensive to measure something in process industry plants. A good example is gas chromatography, which is quite expensive and very sensitive. They break quite often, especially in adverse conditions. In that case, you should calculate instead of measure. So can you calculate it based on the other variables you can easily and cheaply measure? This is where a soft sensor would come into play.”
looking towards the future
dr. Bangert says he's seen a real revolution in the industry in the last 1.5 to two years. “Virtually every company in the chemical, power generation or oil and gas sectors has created a new business department that deals with digitalization, Industry 4.0, IoT, etc. These departments oversee the introduction of new methods in the company”. These departments have two options: either they develop methods themselves and implement them within the company, or they buy external tools and act as consultants to implement the tools. "Most of them first decide to develop their own and then realize relatively quickly that this is time-consuming and expensive," says Dr. Bangert. “By buying an external tool, you can use it today, which gives you a time head start of maybe three or four years and you can focus on your core business. Also, a software company has many customers and receives feedback from many of them, so the software becomes better and more mature much faster than if it were developed in-house.” to expand your business. "In addition to our home market in Central Europe, we are focusing on the United States," says Dr. Bangert, "where we have been active for about a year and a half. In Germany, I would like to get to the corporate deployment phase soon”.
What is machine learning?
Artificial intelligence is primarily concerned with natural languages, such as having your computer give you product advice as you browse a website. dr. Bangert explains: “Machine learning, however, deals with numerical data. For example, if you have 20,000 sensors installed in a refinery and you want to develop a formula that governs the dynamics of that refinery, that's machine learning. Machine learning deals with numbers, AI with words.” Machine learning is the key to all algorithmic solutions, which basically means that the formula is developed automatically without the involvement of a human expert. For more information on machine learning, see the book "Optimization for Industrial Problems" by Dr. Bangert (Springer Verlag) and at www.algorithmicatechnologies. Or if you have any questions, please contact Dr. Bangert at firstname.lastname@example.org
A ML machine learning predictive maintenance improvement
back to blog
relevant blog posts