In this series of blogs, we will explore how to use Azure Machine Learning (ML) for predictive maintenance. It will be completely commercial free until the very last blog post, where we will sneak in a pointer to our Enterprise Asset Management partner who we work with in our Energy & Natural Resources practice. Whether you use our software or not, you will gain valuable insight into Azure ML based predictive maintenance.
Within maintenance management we deal with different care strategies and categories:
- Corrective maintenance: Maintenance is performed as a result of a breakdown or some kind of reported abnormality.
- Preventive maintenance: Time, operation-count, or calendar based maintenance.
- Condition based maintenance: Maintenance is performed based on the actual or current condition of the asset.
- Predictive Maintenance: Maintenance is scheduled as a result of statistical process control often combined with real-time asset monitoring.
Predictive maintenance can be seen as an advanced form of corrective maintenance, where we perform maintenance because we statistically expect a breakdown or an abnormality. The difference between preventive maintenance and predictive maintenance is that preventive maintenance is most often based on empirical evidence and linear projections on operation-counts or hours-in-use, as opposed to true statistical evidence. The obvious weakness of preventive maintenance is that maintenance will often be either too soon or too late, as it is very seldom we see any kind of true linearity in observed machine breakdowns.
Statistical analysis sounds unexciting, knowing that it is at the heart of predictive maintenance, we probably all feel a little discouraged. Stating statistical problems is however, quite fun, so let’s start and just assume we have some wizard at our disposal, who will take care of the analysis part.
What are some of the infinite statistical questions in the predictive maintenance category?
- Do we have correlations between breakdowns? If we can determine statistically that a breakdown in one area of the production line will later cause a breakdown in another area, then we should issue a predictive work order immediately after the first breakdown, thereby avoiding a potentially costly ripple effect of breakdowns.
- Is there a correlation between vibration levels and breakdowns? Alternatively, machine temperature, RPM’s or…. If this can statistically be determined, then predictive action could be taken if vibrations reach a certain threshold.
- Does the skill level of an electrician affect our mean time between failure (MTBF)? In this case you should probably start to require certain certificates for specific work order types.
- Is there a correlation between certain items we manufacture and breakdowns? Maybe the production scheduling engine should route these items to different production resources.
All these questions are from a maintenance management perspective, very interesting. In addition, the answers are even more interesting!
Statistical analysis sounds amazing! If we just knew how to apply it. This is where Azure ML comes to the rescue. The only catch is that in order to get all the Azure ML machinery up-and-running we still need to understand some aspects of statistical tests.
Providing that understanding will be the topic of the upcoming part 2 in this series of blog posts. Stay tuned!