Though machines are prone to break down in the long run, unplanned equipment downtime can throw spanner into production works and heap losses on organizations. With better equipment utilization sense prevailing, reactive maintenance is now giving way for preventive maintenance. As IoT and sensors create another industrial wave, organizations seeking to ride this wave to optimize equipment performance, cut operational costs, and improve productivity also want to acquire the ability to predict asset breakdowns and reduce downtimes.
As organizations seek to leverage preventive maintenance in increasing machine lifetime, predictive maintenance (PdM) models leveraging historical events are giving organizations the much needed agility to predict machine and component failures, take corrective actions to increase uptime of machines and productivity.
Where predictive maintenance model augments preventive maintenance, it takes a classification or a regression route to predict machine failures. Using the right machine learning approach then becomes the prerequisite to achieve desired results from predictive maintenance.
What to consider to build predictive maintenance model?
The modelling strategy is devised by finding answers to queries impacting the model development.
- What is the expected output?
- Is there a need for static as well as dynamic data?
- Are there labelled events at hand to draw a line between successful functioning and failures?
- How many types of failures to be modelled?
- What is the timeframe in question to model a failure?
- What is the performance criteria to be targeted – accuracy, precision or sensitivity?
Which modelling strategy to choose?
Taking the predictive approach to support predictive maintenance calls for the right machine learning modelling strategy. Choosing the predictive maintenance modelling strategy depends on what the model aims to predict and the data available for modelling purpose.
When static characteristics of a machine become the pointer to behaviors, it lays the course to predict how long or for how many days the machine will stay fit and useful. This becomes a regression problem to solve given the possibilities to model the failure path from the historical and static data. More importantly labelled data is a requisite to identify events that falls into a specific type of failure.
Now when the thought ‘when will the machine fail’ becomes the moot question to answer, predictive maintenance modelling strategy takes a classification route. Through this, multiple system failures can be predicted for a defined time window.
Going beyond regression and classification modelling, there are instances when there is no data to support what’s the normal behavior of the machine and what accounts for the machine failure. The strategy here is to build anomaly detection model which detects machine failures without any prior knowledge of those failures.
What modelling technique to use?
PdM cases can be addressed through relevant modelling techniques, which depend on the problem that we are trying to solve and the data at disposal. Here are some scenarios that provide insight into the right modelling technique that can help reap desired results out of preventive maintenance.
Failure within a time-period?
Let’s us consider the scenario of an organization wanting to predict the failure of machine component within a time-period. To acquire the lead time to keep proactive maintenance measures in place, binary classification is the right technique to support preventive maintenance. For using binary classification technique, what becomes a requisite is the training data for identifying failure as well as normal operations.
Failure at this time & what’s the cause
This is a predictive maintenance scenario wherein there is a need to predict if a machine component will fail at a specific period of time and to predict the cause behind the failure. In such cases Multi-class classification rises in relevance. This comes handy for the maintenance team in the way it empowers them look for failure signs during a specific period and take the right action to fix failures.
Machines can fail owing to multiple reasons and hence predictive maintenance turns into a multiclass classification problem – classification because it involves categorization of several failure outcomes as well as reasons inducing machine failure. There are various parameters that are taken into account as that of the health features, risk levels and other reasons for the failure.
Finding patterns to augment predictive maintenance offers another approach. As time-series data becomes a critical component here and when the time period stretches over a long period of time, Long Short Term Memory Networks (LSTM) turns into a good fit in the way they can detect failure patterns over longer time periods. With the problem scenario covering preventive maintenance of machines, deep learning via LSTM networks is an advantageous proposition in the way the networks pick the right features impacting the failure of machines.
The right machine learning approach for predictive maintenance hinges on the problem we are trying to solve, data available to build prediction models and predictions and insights that ought to be delivered through this preventive maintenance exercise.