Have you ever been in a rush getting to work and having to drop off the kids, and by the time you get to the car and you’ve got the kids strapped in you turn the ignition key, but nothing happens. The car makes some weird noises, but the engine won’t start. Now you’ve got to cancel the meeting with the business owners and that big presentation that you were supposed to give because you won’t make it in time – people will be annoyed, because their valuable time will be wasted. Furthermore, the kids will be late for school and you have to pay for an expensive trip by cab. At the end of the day, after having had your car towed to a repair shop, the mechanic informs you that it will be 3 more days before he can get the needed, very expensive, spare parts to repair your car because the fault had caused damage to other parts of the engine.
Now imagine, if you knew 30 days in advance, that a fault was slowly developing in your car’s ignition system, it would be easy to schedule an appointment at a repair shop, plan it into your calendar and make ends meet. You also wouldn’t have to pay for that extra cab ride, because you could simply work from home the day of the scheduled repair, and the kids would not be late for school. As a final added bonus, the fault could be repaired without use of additional, expensive, spare parts because the fault hadn’t come out of control and caused damage to other parts of your car’s engine. All in all, a total win on all accounts; economically, time wise and definitely on the levels of mental frustration. This is what this project aims to achieve, although in a slightly different and bigger setting: power plants.
The ability to detect faults well before they happen using data from the machine in question (vibrations, sounds, temperatures, etc.) is called Predictive Maintenance (PdM), and is what this project aims to improve upon. Specifically, in a power plant setting. Here, many different and critical components exist (pumps, turbines, motors, etc.) that are very expensive to have repaired out of schedule. This problem can be alleviated by continuously monitoring the machines using data collected from them and running advanced machine learning algorithms to detect anomalies in the data, indicating that machines are about to need repair.
This project seeks to address shortcomings to three key areas of power plant PdM. First, how machine learning models can be rolled out to many components that are very different in operating patterns without excessive time spent on customization for each. Secondly, a standard problem to PdM systems is too many false alarms. This causes plant operators to lose confidence in the system and ultimately make them not act on the alarms the system produces. Thus, the project will investigate methods to balance the number of false alarms, while still detecting the true faults. Thirdly, plant engineers have been maintaining these components for many years and have valuable experience and knowledge of them. A driving hypothesis behind the project, is that this knowledge will be useful for continuously monitoring components, if we can incorporate it into the data-driven models. The project will therefore also examine how this knowledge can be used optimally for the prediction of developing faults.
Better monitoring of power plant components will ultimately benefit in terms of vastly lower expenses used on untimely repairs and ordering of key spare parts, a safer and better scheduled working environment for plant operators, as well as ensuring a more stable power and central heating grid for citizens.