Can you predict bearing failures or RUL?

I think that being able to predict bearing failures is something that we are all interested in. At least, it is definitively interesting for those who deal with rotating machines daily. Is it possible? Many people are still working to figure it out, and let me tell you that they are very close to finding more accurate answers. To this point, what researchers have found works better in certain cases than in others. In this article, I will simplify the best I can, how is the general process to estimate the remaining useful life (RUL) of a bearing.

First, let's make a very clear distinction between what the condition monitoring experts do and what is understood by a bearing's RUL estimation or prediction. Condition monitoring is the task of keeping track of machine deterioration as it has occurred. Some would argue that I should use a standardized definition according to ISO or some other standard's institution. But for the sake of simplicity, let's say that this is the main goal of a condition monitoring routine. How frequent you check your bearings depends on how deteriorated they are or, in other words, if you have hit the point where a defect has manifested, and you need to monitor it closely. Take a look at the following curve:

In the figure, condition monitoring frequency increases when the bearing has passed point A. Failure might come so fast in some cases that you won't have time to react. This image is taken from the conference paper: Applications of Artificial Intelligence on Prognostics of Rotating Machinery by Erfan Ahadi and Mostafa Larky, 2018.

If everything goes well, a bearing will not show signs of deterioration until a stage where the defect is detectable. You and I will agree that a bearing starts deteriorating from the moment it is put into operation. The most difficult part of predicting a bearing's life is that you are not noticing these signs of deterioration. Besides, the remaining useful life (RUL) depends on future operating conditions and loads profiles, which we are not aware of yet. You know that anything can happen.

To emphasize it again, RUL means determining the life of a bearing starting from the current moment, even though a defect has not been detected. You can now imagine how complex is this prediction process. Graphically, RUL timing could be drawn like in the next figure.

The figure is taken from maintenance.org - what-is-remaining-useful-life

Taking as a premise that the bearing always deteriorates under their designed regime of operation, researchers created an index to keep track of this deleterious effect. They call it the Health Indicator (HI). Here is an example of the behavior of a HI curve derived from raw measurements.

From raw vibration (upper graph) to HI (lower graph). In orange, the RMS value of the vibration and the HI in blue. The figure is taken from Bach et. al. A Reliable Health Indicator for Fault Prognosis of Bearings Sensors, 2018.

One of the characteristics of this HI curve is that it is monotonically increasing. That is a curve that is always increasing or remaining constant. So for every second the bearing is running, we are damaging the bearing. The HI keeps track of this degradation. But this is not enough to make the magical prediction. We need to find a point where we can "start really predicting." Researchers refer to this point as the time-to-start prediction (TSP). TSP is the point at which the bearing departs from its normal behavior and starts showing evident signs of degradation. The TSP is defined by a percentage of increase of the HI. In other words, when HI hits that level, it is expected that failure will be coming "soon." But, isn't this the same as what condition monitoring guys do? Yes and no. As seen, they measure the current condition of the bearing and compare it with the baseline to see how far they are in the degradation process, but this information is not enough to "predict" the moment of failure. As the defect advances, condition monitoring guys control the machine more frequently to portray how fast it is deteriorating. However, in RUL's case, you aim to predict how soon the HI will reach the TSP using data-driven models; then, you could estimate when the bearing would reach its End Of Life (EOL). That is, you measure how far ahead in the future is the failure. The prediction procedure follows the figure presented next.

Increase of the HI until it hits the limit to TSP. The figure is taken from Wasim et. al. A reliable technique for remaining useful life estimation of rolling element bearings using dynamic regression models from Reliability Engineering and System Safety, 2019.

Now, what should I measure to construct the HI? This is the key question researchers have been devoted to answering. In short, one of the most popular variables to measure is vibration. But it is not as simple as that. Depending on the application, operating conditions, environment, some features extracted from raw vibration values describe the failure better than others. For example, you could use the vibration's RMS, Kurtosis, spectral analysis, or wavelet transformations. And not only limited to these, because there is a broad list of features listed in the literature and even combinations of operations of the previously mentioned features that could be tested.

Researchers have an advantage when studying RUL. They already know the instant of the failure as they are working with historical data. This means they can test, on documented events, the accuracy of their predictions. However, when it comes to real industrial applications, results might not look as promising.

Do you know where researchers get their test data? Continuous vibration data is not often collected from real applications. We don't continuously monitor equipment as I would dream of doing it. We don't have that many 24/7/365 vibration monitored machines. But, and if we did, it would take years from the moment a bearing is installed until its failure, hopefully. It would be great to have this information, definitively. The data is there, but we need to harvest it.

The workaround to this problem of data availability and slowness of results (failures) is to accelerate failure in test bearings and use this data for research purposes. This means no real equipment is hurt in the process. There are two popular data sets, the Intelligent Machine Systems (IMS) NASA and the PRONOSTIA dataset. The latter dataset comes from a machine whose sole purpose is to destroy bearings while monitoring the whole process. Bearing failure is usually reached in a matter of hours instead of years. You can now imagine the wealth of data obtained from this source. Take a look at how the machine looks like.

Taken from Nectoux et. al. PRONOSTIA: An experimental platform for bearings accelerated degradation tests. Conference on Prognostics and Health Management, 2012.

To wrap this up a little bit. If you continuously could measure a condition parameter from your machine, usually vibration, and select a feature from the raw data, you could build an HI and model how this HI will behave in time to estimate the RUL of your bearings. You repeat the process until you find the right feature and model to match the failure. Remember, some features would describe better than others bearing failure in the target machine.

Due to the inconveniences mentioned here and more, we know that this is not an easy task. Nor you could say that one prediction model fits all. But, if you are an OEM, you have a tremendous advantage with respect to generic applications. You can develop a model able to read the failure signatures of your target equipment, your equipment. This would give you some sense of repeatability in your predictions' success, ending with an RUL prediction model that fits your product. Maybe ??

Finally, no matter what, we need to collect data and better if it is in continuous form. Take a look at my article Where is my data? You'll find a couple of interesting cases there. Thanks!