Select Page

In Part of this series (Jan. 16, 2020, link below), we discussed a few basics of supervised machine-learning (ML) systems, including typical steps to the development of an ML solution. In this second installment, we take a brief look at the types of algorithms used in these types of system models, which primarily relate to statistical processes.

Overall, the concepts surrounding the use of machine learning and predictive analytics in the RAM arena are very similar to much of what we’ve already been doing. The significant difference with the use of ML is to provide the ability to manage large amounts of data using tools that would be used to identify patterns, identify faults, or project Time to Failure Estimation (TTFE), which is now sometimes referred to as “Remaining Useful Life” (RUL). (Reference:  for the original IEEE 2009 paper introducing this topic.)

To provide accurate estimates, which are expected to become more accurate as the fault is tracked in continuous predictive models, real or simulated failure data is normally used. The simplest form of analysis from this point is curve-fitting failure data and determining TTFE from the mean average or, for alarms, verifying the curve for normal operation, setting failure bands, and tracking when data crosses those bounds. This would normally include the items we most frequently observe in traditional predictive maintenance, which are upper and lower control limits.

Another method (which we will not cover) is when neither failure data nor a model exist. In that case, you can use unsupervised methods, including clustering, to review probabilities and ascertain where outliers from real data represent potential issues based upon some time-based measurement and best guess (data science).

One frequently used method for projecting the condition of equipment and TTFE involves regression-analysis models, the most common being linear regression. In a regression model, the independent variable may be time, number of operations, etc., and the dependent variable relates to the measurements taken to determine some outcome.  In a good linear regression model, as components approach failure, or age, or some other condition that is being measured for, a slope should exist when a line is drawn through the average of the datapoints.  The point in time or operations that the line intersects projected failure values would be your projected TTFE.

Depending on the type of equipment and/or materials you’re dealing with, the regression model may be more accurate as a non-linear model. A common application for non-linear regression models may be found with modern dielectric materials (electrical insulation), in which projected life is non-linear in relation to temperature and resistance to failure over time (non-linear regression). Bearing and mechanical systems tend to be more linear in nature (linear regression).

The work in projecting life related to electric machines includes a combination of TTFE for insulation systems in a variety of environments IEEE 2014), plus an evaluation of mechanical components. This would identify that a machine-learning evaluation of an electric motor or generator outside of just mechanical components, such as the bearings, requires a combination of regression modeling or the selection of another solution for TTFE.  That’s one of the reasons why vibration-analysis ML models are more common than those designed around electrical insulation failures. In fact, insulation-failure models had been considered inaccurate in lower-voltage systems until publication of the 2014 IEEE paper that was part of a study on projecting insulation and mechanical life of hybrid-vehicle motor designs for a manufacturer.

As we pass through the number of potential regression methods that are primarily dependent upon the number of observations used to build a successful model, we fall into solutions that would fit a complex electric-machinery model. This includes such things as regression-tree models, which provide a weighted combination of regression models. As a result, we can have both mechanical components (linear) and electrical insulation components (non-linear) built into the model. Even in the more common regression methods, with enough information, we are able to build relatively accurate models that can be used along with continuous-monitoring systems or a combination of data inputs, including manual data collection and sensors to make sound decisions.

In our next article, we’ll discuss some of the more complex models that can be used to increase confidence and accuracy in machine learning.TRR