How does Predictive Maintenance Work?
by Frederick Mannings, Data Engineering Lead
Predictive maintenance algorithms work by collecting data from your system, constructing conditional indicators from this raw data, and feeding these conditional indicators to Remaining Useful Life (RUL) models that are then able to predict how long a system can continue to operate before failing.
Conditional Indicators
After collecting the system’s data, the next step is to transform this raw data in order to create conditional indicators from it. This step can be referred to as feature engineering. The raw data is filtered, encoded, and new variables are constructed using a set of techniques including Fast Fourier Transform (FFT), polynomial features, and physics-based models, just to cite a few. The objective is to end up with variables that strongly correlate with the time to failure of the system, and this is the reason why they can be called conditional indicators – simply because they represent the condition of the system. In practice, it is common to create multiple conditional indicators, one for each sub-system.
Taking the example of an industrial robot, it would be good practice to create a conditional indicator to monitor battery health, another one for the gearing condition of each joint, and so for each sub-systems of the robot. All of that makes the creation of conditional indicators a difficult activity that requires a lot of experience in the fields of engineering , machine learning, and signal processing. At Predixus, we have specialised at developing methods and techniques based on our experience to make this step as efficient and successful as possible.
Remaining Useful Life
After processing the raw data to create conditional indicators, the last step of predictive maintenance is to use these indicators to predict the Remaining Useful Life (RUL) of your system. This is made possible by using time-series analysis to estimate the time it will take for the conditional indicator to reach a critical threshold, or by using machine learning to learn the time-to-failure of your system by looking at the data from the past. However, the goal here is not only to predict the remaining useful life of your equipment, but also the uncertainty around this number.
For instance, we want to inform the engineers that a machine is most likely to fail in 1000 cycles, plus or minus 100 cycles, with a confidence interval of 98%. This means that if you wait 900 cycles before organising maintenance, there would be only a 1% chance to get an early failure on that machine. This confidence interval is usually updated every cycle and narrows down as you get closer to real system failure, which makes it almost impossible for predictive maintenance to missout a failure. All in all, using this information in an Asset Management System gives you the ability to optimise your maintenance strategy, manage your risks, track your maintenance costs, whilst simultaneously handling thousands of systems.