How to Use Data in a Predictive Maintenance Strategy

How can engineers and data scientists get ROI from their predictive maintenance program? In this blog, we’ll run through some of the potential uses of industrial data, and how the resulting industrial analytics form part of a cohesive maintenance strategy.

If you missed the first part, where we covered the types of data that are useful in a predictive maintenance program, you can check it out here.

Now that pain points are known and outlined, valuable or bad actor assets identified, and the state of data has been captured, industrial businesses have the context to understand — and realize — the impact of predictive maintenance and reliability strategies.

First, industrial businesses need to understand the potential uses of their data. Questions might come up like:

• What do you expect to achieve from a targeted transition to predictive reliability?
  
• Do you have relevant root causes for equipment failures?
  
• What failures would be most beneficial to predict, and at what lead times?
  

There are four major benefits from a transition to predictive maintenance. The first two benefits are not truly predictive in nature, but they are required for predictive analytics.

In turn, they are often bundled into a software solution but are broken out here.

This is sometimes called an equipment dashboard or KPI dashboard and is simply a way of providing the user with an intuitive way of visualizing the equipment data. This can take a format that drives problem recognition and programmatic efficiency.

It is not data analytics per se, and does not run the data through any sort of AI or analytics algorithms. The data ingestion pipelines required to perform data modeling are often first surfaced onto a dashboard as a way to verify, validate, and adjust the data connectors that will be used later.

For example, a data surfacing dashboard could provide a screen that shows all trucks in a given fleet that will require preventive maintenance in the next week, like the below.

The preventive maintenance schedules would be ingested from existing Work Management System (WMS) software in a read-only type schema. This information would be available to dispatchers during route assignment activities without having to open the WMS preventive maintenance screen for each truck.

It's one of the reasons why shared data access is a core part of the team-first approach to Industry 4.0 advances like predictive analytics.

Free-Text and label correction engines are a solution to clean up missing or inconsistent work order and parts order data. Pattern recognition algorithms can replace missing items such as funding center codes. They also fix work order (WO) descriptions to match the work actually performed. This can often yield a 15% shift in root cause binning over non-corrected WO and parts data.

An example would be a WO that was originally written to repair a pump, but upon closer investigation and troubleshooting, the maintenance crew determined that the problem was caused by a faulty motor starter that required replacement.

By training the "cleanup engine" to look at the resources actually used for the repairs (electricians instead of mechanics) as well as the parts used (a motor starter and not mechanical pieces or parts), the WO description can be changed to reflect the true issue being addressed. This data correction improves future trend reporting.

With programmable logic controller-generated threshold alarms (like an alarm that is generated when a single sensor exceeds a static value), “nuisance” alarms are often generated and then ignored.

These false alarms quickly degrade the culture of an operating staff as their focus is shifted away from finding the underlying problem that is causing the alarm. In time, these distractions threaten the health of the equipment, as teams focus on making the alarm stop rather than addressing the issue.

A multivariant approach "cuts through the noise" by finding the trends and patterns behind the seemingly random alarm chatter.

This approach should allow operators to focus on what is truly important. By running AI algorithms on these single-point alarms, patterns emerge that improve both the lead-time of true positives as well as by greatly reducing the number of false positives.

Operators are provided insights that are:

1. in advance of the problem,
2. can be trusted to be real, and
   
3. are specific enough to allow the alert to be isolated to one machine or subsystem and to a small group of related failure modes.

This is the bread and butter of data analytics software — the area of cool AI that uses multiple variables related to a machine to statistically predict a given failure mode.

Implemented correctly, these models are truly amazing in that they can often predict an imminent failure sometimes days in advance, and from seemingly unrelated data sources. Predictive insights allow time for an organization to prepare for an equipment downtime as opposed to reacting to an emergent failure.

It is unrealistic, however, that there will be an analytics model for every potential failure mode for a given piece of equipment. It then makes sense that any remaining failure modes (those not being detected by the analytics model) would be continued to be addressed by more conventional methods (performance of a time-based or condition-based maintenance task).

With the fraction of industrial assets that wear out in predictable ways by time and usage alone, it still makes sense to stick to preventative maintenance in some instances.

Any comprehensive maintenance strategy or long-term asset management plan must start with a thorough understanding of the failure and degradation modes, locations, applicable stressors, and elapsed time for the first detection of that failure or degradation.

An understanding of the impact of maintenance being performed on short and long-term reliability is required to properly understand the health of equipment. Each cause and location of equipment failure should be understood in relation to the amount of time before that cause is detectable for that location, and any maintenance tasks being performed should be mapped to the degradation modes that it mitigates.

As a result, the maintenance task being performed is the mitigation strategy for this degradation. This is equally true when the mitigation strategy for a given failure mode is the implementation of a predictive analytics model. The output of each predictive analytics model should be mapped to the failure modes or locations that it mitigates. This failure mapping allows for a comprehensive understanding of the total maintenance strategy as it relates to the mitigation of applicable failure modes.

For any degradation that is completely mitigated by a predictive analytics model, conventional maintenance tasks for the same degradation can then be eliminated. If the task also mitigates degradation modes outside of those bounded by the predictive analytics model, it may be appropriate to adjust its scope and to also adjust the performance frequency.

Predictive analytics is an exciting and promising way to improve the reliability of key assets and components. Moving to a predictive maintenance program doesn’t happen overnight, and even once in place, preventive maintenance tasks may still be appropriate for some types of industrial equipment. Predictive maintenance cannot be done in a vacuum.

Engineers and industrial data scientists need a thorough understanding of the failure modes and locations being detected by the predictive analytics model as well as by the “legacy” PM tasks. Once they do, they’ll be able to tap into the great value of advanced warning on their assets.

Lucinda Reynolds, CMRP is a Manager of Reliability Engineering at Uptake and has over 30 years of experience in cross-industry experience in machine maintenance and reliability.

Ready to put your data to work?