How to Scope Data Use for a Predictive Maintenance Strategy

The areas of Artificial Intelligence (AI), Machine Learning (ML), Industrial Data Analytics, and Industrial Internet of Things (IIoT) seem to be the latest ‘Last Frontier’ of equipment reliability. There is an attitude out there that contends that the adding of sensors will automatically result in an improvement in reliability.

I hate to be the bearer of bad news, but hooking sensor wires up to a piece of machinery will not automatically give it increased reliability, better breath, or any other often-advertised superpowers. That being said, there are some very real benefits (and yes, cost savings) that can be realized by a disciplined implementation of advanced industrial analytics into an existing reliability program.

The usefulness of AI/ML (as well as predictive maintenance (PdM) and preventive maintenance (PM)) is not just to prevent failures, but to give you an accurate assessment of the measured (or inspected) parameters.

In other words, the AI piece of the puzzle is drawing data from sensors and other data sources and distilling this information into something that you, as a human, can understand and act upon. The resultant decisions and actions are what prevents failures.

So, when you think about it, the goal of all of these acronyms is to provide accurate, understandable information available in a timely manner so that you can make those decisions as quickly and accurately as possible. Due to the lack of maturity in the marketplace (both on the vendor-side as well as on the consumer-side), most organizations don’t know how to distinguish the several offerings that are out there.

As a result, millions could be spent based on the charisma of the salesperson and nothing else. Creating a smarter consumer that understands where their organization is currently, where they want to go and the various options available on that journey is the goal of this blog post.

Below is a list of some things to consider and questions to ask when considering or establishing a predictive maintenance project plan.

This should include an unflinching assessment of which machines are bad actors. These are critical assets that are responsible for lost revenue and higher than necessary overall costs. This evaluation includes:

• Preventive Maintenance expenditures

• Corrective Maintenance expenditures

• Capital (Modification) expenditures

• Costs of lost production

• Costs of scrapped product

• Safety, Environmental and Regulatory cost impacts

Any project plan should rank components by the above criteria and prioritize steps and components according to these criteria. The transition to predictive maintenance should be performed on the most impactful component type first, and then, when the process and pitfalls are well understood, rolled over to other components in order of importance and return on investment.

Nobody’s data is perfect, but here are some things to consider before you start out. There are generally three different classes of data.

Time-Series Data (or data that indicates the current state of an asset over time)

• Live Sensor Data
  • Are the most important components sensored?
    
  • Are those sensors, alone or in combination, able to detect degradation for your most troublesome failure modes? Again, just sensoring components does not inherently improve performance.
    
  • How exportable or accessible are these sensor readings (do they stay in the programmable logic controller (PLC), are they already being sent to a data lake or estate, etc.)?
    
  • What is the cadence or sampling rate of the sensor readings?
    
    • Sampling rates may have to be adjusted based on specific failure modes being detected and on the desired lead-time for these failure modes. You are not going to get a half-hour lead-time on a prediction when your sampling rate is at once per hour.
      
  • Are there currently PLC or controller logic threshold alarms for the equipment?
    
  • Are these alarms retrievable either in live-time or near live-time?
    • These are not predictive in nature but can be used to establish predictive data models.

• Inspection Reports (alarm response inspections, RCI follow-up inspections, post-maintenance testing)
  
  • Are these records digitized?
    
  • Are these records indexed in a way that provides for easy retrieval (by date, by equipment ID number, etc.)
    • Indexing an inspection to an RCI document number may make sense at the time, but ultimately, it needs to be retrievable as part of the equipment or history record of each applicable component.

• Rounds sheets (thermography, vibration, etc)
  • Are these records digitized?
    
  • Are these records indexed in a way that provides for easy retrieval (by date, by equipment ID number, etc.)?
    • Indexing an inspection to a periodic inspection activity may make sense at the time, an index to the equipment is still required for proper assessment of each component’s individual health.

Data that tells us what was done to assets

• Historical Work Order and Parts Data
  • Is there Work Order history data that is consistent and accessible? Where possible, greater than 5 years of historical data is valuable.
    
  • Work order history should include:
    • As-found condition (either codes or text)
    • A description of work performed
      
    • Resource hours expended
      
    • Parts replaced (noun name, quantity, and price)
  • Free-text and cost center information should be ‘cleaned’ by a supervised learning free-text cleaning program to ensure that any trends being shown by the data are accurate.

• Operator Logs and Shift Notes
  • Are these records digitized?
    
  • Are these records indexed in a way that provides for easy retrieval? Indexing by date will usually be the only option here, but cross-indexing should be performed as well.
  • Consideration should be given to ‘cleaning’ free-text formation to allow for correlation (indexing) and subsequent trending of data at the individual component level.
    

• Lubrication Records
  • Are these records digitized?
    
  • Are these records indexed in a way that provides for easy retrieval (by date, by equipment ID number, etc.)?
    • Indexing lubrication to a periodic activity may make sense, but a cross-reference to the equipment is still required for proper assessment of each component’s individual health.
      
  • Consideration should be given to cleaning free-text information to allow for correlation (indexing) and subsequent trending of lubrication history at the individual component level.

Data that tells us about Operational Experience that are applicable to given components

This can be data like asset failures, RCIs, modifications, external industry notices.

• Design Changes (Modifications)
  • Are these records digitized?
    
  • Are these records indexed in a way that provides for easy retrieval (by equipment ID number, etc.)?
    
  • Are these records also cross-referenced to any ‘driving’ documents (RCI, corrective work orders, external notice, etc.)?
    

• RCI Determinations
  • Are these records digitized?
  • Are these records indexed in a way that provides for easy retrieval (by date, by equipment ID number, etc.)?
  • Are these records also cross-referenced to any ‘driving’ or corrective actions documents (corrective work orders, external notice, design changes, etc.)?

Don’t worry if your organization doesn’t know the answers to each of the above questions — and don’t worry if you do know the answers but don’t like them.

Industrial intelligence should start with where you are today and should be able to establish a clear roadmap to where you want to end up. Those questions above are intended to prompt questions like:

1. Is my process data sufficient to know the current state of my pumps?

2. Do I have any records of what actually caused my past turbine failures?

Once you have run down this list of questions, you should have a good idea of where to start with your data for it to be ready for analytics and predictive maintenance.

It may take time to scope out the potential uses of industrial data, and how to create a strategy for unified data management to make these use cases possible. In part II of this blog, we’ll cover how industrial businesses can begin to realize the value of their data.

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?