Approaches to Predictive Maintenance
For predictive maintenance to succeed, there are three main aspects that must be present.
First, and probably foremost, you need quality data. Ideally, you want historical data that takes into account events that have, in the past, ended in failure. Failure data needs to be juxtaposed against static features of the machine itself, including its average use, general properties, and the conditions under which it operates.
You will no doubt end up with a lot of data, so it is critical to focus on the right data. Getting hung up on extraneous information does little more than muddy the waters, deflecting attention away from what’s most important. You should ask yourself, what failures are likely to occur? Which ones do you want to predict? What happens when a process fails? Does it happen fast, or is it a slow burn over time?
Finally, take a close look at any other related systems and parts. Are there other components that are related to the failure? Can their performance be measured? And finally, how often do these measurements need to happen?
Data collection needs to take place over an extended period for best results. Quality data results in a more accurate predictive model. Anything less will only narrow the field of possibilities rather than give you hard truths. Analyze the available data and ask yourself if it is possible to build a predictive model based on these insights.
It is important to have the proper context when looking at a problem, as only then do we have the ability to evaluate the predictions with some accuracy.
In general, we use one of two predictive modeling approaches:
Regression Models predict the remaining useful lifetime of a component. It tells us how much time we have left before the machine fails. For a regression model to work, historical data is necessary. Every event is tracked and, ideally, various types of failure are represented.
The assumption offered by the regression model is that, based on the inherent (static) aspects of the system and its performance in the present, its remaining lifecycle is predictable. However, if there are several ways in which a system can fail, a separate model must be created for each possibility.
Classification models predict machine failure within a certain window of time. In this scenario, we don’t need to know too far in advance when or if a machine is going to fail, only that failure is imminent.
Classification and regression models are similar in many ways, but they do differ on a few points. First, the classification looks at a window of time rather than an exact time. This means that the gradation of the degradation process is a little more relaxed, requiring less exacting data.
Additionally, the classification model supports multiple types of failure, allowing incidents to be grouped together under the same classification. The success of a classification model depends on there being enough data available, and enough instances of certain types of failures to inform the ML model.
How It Works
Once modeled, predictive maintenance proceeds in this way:
The ML model collects sensor data and, based on historical failure data, identifies the events that precede a failure.
We pre-set the desired parameters to trigger an alert to a potential failure. When the sensor data breaches these parameters, an alert is initiated.
Where machine learning comes in is in detecting unusual patterns that are outside normal system operation. With better awareness of these anomalies based on quality data, our ability to predict failure improves dramatically.
In conclusion, machine learning supports the analysis of vast amounts of data with minimal human intervention. Applied using best practices, it is the ideal approach to cost reduction and risk mitigation.
By applying machine learning, combined with data collected from IIoT devices, it is possible to improve processes, reduce costs, optimize employee efficiency, and reduce machine downtime significantly – all critical aspects of a successful organization.
If you would like to learn more about machine learning and how it can support a predictive maintenance posture, reach out today.