The Essential Guide to

Predictive
Maintenance
Applying Splunk and machine learning to your
maintenance operations
Compete and Save Table of Contents
Introduction...................................................................................................3

With Predictive Let’s Break It Down.....................................................................................4

Planned maintenance vs. unplanned maintenance..............................................4

Maintenance
Shifting from preventive to predictive maintenance..........................................4

The Solution Defined..................................................................................5

Want to See How You Can Get Started?...............................................6
Splunk’s Stages to Predictive Maintenance........................................6
3 Major Types of Predictive Maintenance............................................7
Initial product defect and recall analysis....................................................................7

Unplanned maintenance by early detection.............................................................7

Recurring maintenance optimization............................................................................7

Getting Specific............................................................................................8
Use Case 1: Initial Product Defect and Recall Analysis......................................8

Use Case 2: Unplanned Maintenance and Early Detection..............................9

Use Case 3: Recurring Maintenance Optimization............................................. 10

Want to Know With the ongoing explosion of inexpensive sensors and accessible
big data technology, the way industrial organizations look at assets is
completely changing. Critical, forward-leaning analysis of operations

How Predictive is no longer a matter of “what if” but of what can we learn, when and
how fast. This is a boon for industrial operations, where opportunities

Capabilities that leverage operational technology (OT) data have the potential to
disrupt overall bottom lines in just about every industry.

Will Change This is especially true of asset performance, as organizations shift

from preventive to predictive maintenance. What company doesn’t

Maintenance? want to be able to analyze data and predict when it’s the optimal time
to do maintenance or predict failures of critical operational assets?
It can lower costs, boost profits and improve the overall efficiency of
the business.

By following Splunk’s four-stage journey, you’ll be doing predictive

maintenance in no time.

3
Let’s Break It Down Unplanned downtime results in loss of customer trust
and productivity — 46% couldn’t deliver services to
customers, 37% lost production time on a critical asset,
The ultimate goal for predictive maintenance is to optimize when
and 29% were totally unable to service or support
and how often maintenance is performed on industrial assets
specific equipment or assets. 2
with data. Armed with the ability to discover patterns and signals
from sensor data, organizations can look around corners, apply
Forward-looking organizations are now seeking to more accurately
timely maintenance strategies, and ultimately predict and plan
predict problems instead of simply preventing them with
for big future issues that cause massive downtime (and massive
standardized maintenance. It’s about time too: 70% of companies
maintenance costs).
are unaware of when equipment assets are due for maintenance
or upgrade. 3
Planned maintenance vs.
unplanned maintenance Shifting from preventive
Most organizations are concerned with two types of maintenance —
planned and unplanned.
to predictive maintenance
Preventive maintenance is planned maintenance to preempt
Planned or scheduled maintenance happens before a failure failures caused by wear and tear. It’s often expensive and not always
occurs — or prevents failures from happening. Unplanned necessary, although it is currently the safest option. The eventual
maintenance, also known as reactive maintenance, happens when goal — predictive maintenance — is also planned maintenance, but
parts fail and need replacing. This type of maintenance is generally it takes into account estimated service intervals as well as data-
costly, as it affects operations due to asset availability, kinks in the driven insights based on the measurement of operating conditions,
supply chain and more depending on your industry. This also often to monitor and diagnose equipment issues in real time. It’s meant to
means business grinds to a halt or slows down — and services are catch anomalies in operations before they become major challenges
interrupted for customers. that could affect the business. Predictive maintenance also makes
the most of asset resources and extends an asset’s life as proper
and timely maintenance becomes more feasible, and eventually,

Of the 82% of companies that have experienced integral to its life cycle. Still, it’s no easy feat. According to a recent
report, 24% of operators describe their maintenance approach as a
unplanned downtime over the past three years, those
“predictive” one, based on data and analytics. The rest either took a
outages lasted an average of four hours and cost
reactive or time-based approach.4
an average of $2 million.1

1. h
 ttp://lp.servicemax.com/Vanson-Bourne-Whitepaper-Unplanned-Downtime-
LP.html?utm_source=blog&utm_campaign=vansonbourne2017

2. h
 ttp://lp.servicemax.com/Vanson-Bourne-Whitepaper-Unplanned-Downtime-
LP.html?utm_source=blog&utm_campaign=vansonbourne2017

3. h
 ttps://www.geoilandgas.com/sites/geog.dev.local/files/ge_offshore_study_paper.pdf

4. h
 ttps://www.geoilandgas.com/sites/geog.dev.local/files/ge_offshore_study_paper.pdf

4
The Solution Defined
Industrial organizations, from manufacturing to oil & gas, need
to be able to ascertain the state of their assets before it’s too
late. Companies need systems that conduct advanced analytics
and machine learning so they can have the insights they need to
predict issues instead of continuing to react to them. Fortunately,
the ecosystem has finally reached the point at which it no longer
has to be reactive or preventative. Condition-based and predictive
maintenance are starting to change the industrial world.

The Splunk platform makes this possible. With all your data in Splunk,
you can easily leverage Splunk machine learning tools to understand
your data and construct machine learning models to start taking
predictive action in no time.

Splunk for Industrial IoT simplifies data preprocessing through a

series of exploration techniques to provide a clear picture of your
assets, what the metrics mean in the context of the assets and their
components, and how the metrics change over time. Armed with the
appropriate techniques for your dataset and the insights they provide,
you’ll avoid costly repairs while maximizing the use and availability of
the equipment.

5
Want to See How Splunk’s Stages
You Can Get Started? to Predictive
Let’s walk through the process of effective predictive maintenance.
We will also explore top use cases, helping you dive into what
Maintenance
predictive maintenance can do in different settings so you can reach
optimal performance. Stage 1:
Use Splunk software to collect,
Data store and structure asset
Collection + metrics
Ingestion

Stage 2:

Pre-processing and data

Data exploration help you understand
Exploration the type of data in use and the
characteristics

Stage 3:

Conduct anomaly detection via

Methods of supervised and unsupervised
Analysis machine learning for predictive
maintenance analysis

Stage 4:

Apply predictive models to

Operationalization broader implementations —
create reports and alerts for
operational actions

6
Next, we will provide you with specific industrial IoT maintenance use
cases with which you can practice each stage. The goal: helping you
understand and effectively navigate each stage in differing use cases
so you can effectively make it a practice in your organization.

3 Major Types Unplanned maintenance by early detection

Summary: Downtime can be costly, and if operators miss the signs of

of Predictive failure, it can leave them scrambling to fix the problems. Early
detection is critical so operations don’t grind to a halt.

Maintenance Unplanned

Initial product defect and recall analysis

Metrics
Summary: Not all products are created equal — whether it’s vehicles,
turbines, machinery or mobile devices — especially when they
are first deployed. It’s important to understand potential points of
failure as these assets make it out onto the field or factory floor.
Time
With connected technology, this has now become feasible and
almost necessary.
Recurring maintenance optimization
Defects Summary: Preventative planned maintenance doesn’t always
mean everything runs efficiently. In fact, it could be the opposite if
you’re using resources on maintenance for assets that don’t need it.
Metrics

Optimizing maintenance cycles through a data analytics approach

ensures you do maintenance at the right time to maximize value and
reduce cost.

Optimal
Time

Metrics

Time

7
Getting Specific How to implement?
Collect any performance metrics related to early life stage — typical
metrics are temperature, pressure, flow rates or vibrations. The goal
Let’s dive deeper into these use cases to get a better grasp of their is to feature engineer the initial performance metrics that represent
importance and how you can leverage them in your operations. initial expected performance, then classify the defects against
what’s normal.
Defects
Methodology summary:
• Conduct analysis for an initial asset performance profile and
understand what the expected initial performance metrics
Metrics

combinations look like.

• Engineer feature characteristics of an initial performance status

from multiple assets.
Time • Some of the engineered features could be average, median, ranges
and quartile analysis of multiple asset samples.
Use Case 1: Initial Product Defect and
Recall Analysis • Identify the initial asset performance that’s positioned outside of
major standard deviation bell curves.
Description
In the example of heavy industry vehicles or any connected cars, One example may be collecting data from a vehicle’s onboard
monitoring for potential defects during operations can create mobile data terminal (MDT) and analyzing it via cluster analysis to
opportunities for the manufacturer to not only monitor for any group certain behavior across a fleet. Externally added sensors
defects but also to personalize maintenance support services. All of in commercial/construction vehicle fleets can also collect data
this reduces vehicle failures and optimizes operations. vibrations, temperature readings and more.

Why is it important? Data sources

Imagine an automobile manufacturer launches a new car model • Vehicle mobile data terminals provide relevant metrics.
but soon finds out there is a potential batch of vehicles that could • External sensors tracking variables like vibrations also provide
be affected by critical engine defects. With data collected from relevant data.
millions of cars, the manufacturer can analyze whether the defect
exists or not. If it does exist, the manufacturer knows exactly which
vehicles need to be recalled. This kind of insight can save millions of
warranty costs.

8
 Unplanned
Metrics

Time

Use Case 2: Unplanned Maintenance and Methodology summary:

Early Detection • Conduct analysis for an asset performance change profile and
better understand the changes in performance metrics over time.
Description
Detecting downtime or equipment failures is another way machine • Calculate the R-squared (R2) between the expected increase rate
learning techniques can be applied to get results. This also applies and the actual increase rate.
to monitoring asset performance metrics at a more granular level. • Identify the spikes in R2 between expected and actual events to
Even if assets seem to have more life left, they can unexpectedly stop define the threshold.
when a component fails. To detect unexpected failures, the analysis
Data sources
looks for signs and patterns that fall outside the norm.
Assets like turbines, motors, engines, compressors and more provide
metrics for analysis.
Why is it important?
A large percentage of maintenance is related to unplanned downtime.
It’s costly, can affect operations downstream, and can damage a
company’s reputation with customers and users.

How to implement?
Collect any performance metrics related to asset life stage as well as
typical metrics like temperature, pressure, flow rates and vibrations.
The goal is to apply deviation analysis on metrics that change more
rapidly with bigger dispersion so as to note any fast failing status for
an asset.

9
 Optimal
Metrics

Time

Use Case 3: Recurring Maintenance Optimization Methodology summary:

• Conduct analysis for an asset performance profile and understand
Description
performance trends/patterns of metrics as the assets wear out.
A pimary directive of an operations manager is to make sure assets
are running at peak performance, while keeping costs under control. • Apply either anomaly detection, unsupervised learning or
This means maintenance done right. Too much maintenance can be supervised learning to find the pattern of performance metrics that
costly and too little maintenance could mean disaster. Optimizing deviates from the norm or expected ranges of operation.
maintenance cycles is key in getting this process right. – Anomaly detection approach:
Detect outliers by applying a standard deviation method
Why is it important?
– Unsupervised learning approach:
Imagine you’re an operations manager for a massive airline. Keeping
Cluster similar behavior performance metrics to group different
your fleet’s jet engines running is a key component to ensuring
conditions of assets
the availability and safety of your service. Recurring maintenance
– Supervised learning approach:
optimization plays an integral role in this process as failing to do it
Design a training model by labeling training data, and then
properly could mean up to $1 million a day in lost profits per jet.
applying the model to future date-sets to be predicted

How to implement?
Data sources examples
Collect data from your fleet’s engines while considering other factors
Jet engine metrics usually consist of 13 major sensors metrics
like flight distances, altitudes and directions. The goal is to apply a
including various speeds, temperatures, pressures and liquid
data-driven, analytic approach to extend the maintenance life-cycle
flow ratios.
policy that would optimize the policy per engine, not for the fleet
as a whole.

10
About Splunk
Splunk Inc. makes data accessible, usable and valuable to everyone.
Download the Splunk Essentials for Predictive Maintenance and get
hands-on experience with machine learning for predictive maintenance.

Splunk, Splunk>, Data-to-Everything, D2E and Turn Data Into

Doing are trademarks and registered trademarks of Splunk Inc.
in the United States and other countries. All other brand names,
product names or trademarks belong to their respective
owners. © 2020 Splunk Inc. All rights reserved.

20-20901-SPLK-Essential-Guide-Predictive-Maintenance-104