SAP Thought Leadership Paper

Predictive Maintenance

Using Predictive Maintenance to Approach

Zero Downtime
How Predictive Analytics Makes This Possible
2015 SAP SE or an SAP affiliate company. All rights reserved.
Table of Contents

4 Optimizing Machine Maintenance

4 Reactive Versus Predictive Maintenance

5 Predictive Modeling

6 Predictive Maintenance Use Cases

6 Transportation

6 Manufacturing and Production

7 Utilities

8 Medical Equipment

8 Data Centers and Clouds

9 Enabling Predictive Maintenance

9 SAP Predictive Analytics

12 Learn More

Paul Pallath, PhD

Chief Data Scientist and Director, SAP Advanced Analytics, Dublin

2015 SAP SE or an SAP affiliate company. All rights reserved.

2/12
Using Predictive Maintenance to Approach Zero Downtime

With the Internet of Things, machine-to-machine

communication, and connected systems, large volumes of
high-velocity data streams are now available that capture
the behavior of machines in real time. Companies can
employ powerful new predictive analytics to perform
modeling that enables predictive maintenance for these
assets, with the goal of achieving zero downtime.

2015 SAP SE or an SAP affiliate company. All rights reserved.

3/12
Using Predictive Maintenance to Approach Zero Downtime

Optimizing Machine Maintenance

REACTIVE VERSUS PREDICTIVE MAINTENANCE With recent advances in the Internet of Things,
Until now, machine maintenance has been either machine-to-machine communication, and con-
reactive (performed when failure occurs) or based nected systems, it is possible to stream sensor
on some heuristic (such as servicing the machine readings from machines so their health can be
every n hours of continuous operation). Parts monitored in real time. Due to the high velocity
that fail earlier than expected can be very expen- and volume of data being streamed, it is not
sive and time consuming for businesses, because humanly possible to check all sensor values. But
the necessary tools, components, and people may advances in predictive analytics have made it
not be available for immediate response result- possible to evaluate the changes in the patterns
ing in machine downtime. But since such incidents of the sensor readings and identify the chances
are sporadic, it is not economically feasible for a of machine failure significantly in advance of the
company to stock all parts and tools and employ event. It is also possible to identify the probability
resources just to have them at the ready; this of other parts failing in due course by analyzing
results in unnecessary costs. To counter this prob- the historical patterns of failures, so that preven-
lem, companies perform scheduled maintenance tive maintenance can be planned. Capturing the
using either statistical concepts like mean time changes in patterns in the sensor feeds in real
to failure or other engineering practices based on time, and using that information to predict a pos-
historical experiences. However, this approach sible failure and schedule a proactive mainte-
has its own challenge: parts can be replaced or nance strategy, is called predictive maintenance.
serviced too frequently, which drives up costs
unnecessarily. There are also instances where
machines fail before the scheduled planned
maintenance due to other reasons.

With the Internet of Things, machine-to-machine

communication, and connected systems, sensor
readings from machines can be streamed so their
health can be monitored in real time.

2015 SAP SE or an SAP affiliate company. All rights reserved.

4/12
Using Predictive Maintenance to Approach Zero Downtime

PREDICTIVE MODELING cost functions are, classifiers or models might

Modeling for accurately predicting the probability completely ignore the rare event. Therefore,
of occurrences of rare events like machine fail- special attention should be given to the choice
ures events that occur very infrequently and of performance metric being used in general
are very hard to detect requires robust algo- and, more important, for cases with unbalanced
rithms. Rare though they are, these events are classes (where one class of responses is dispro-
normally associated with very high costs, so portionately larger than the opposite class).
efforts to predict them are necessary. There are
two related aspects that make building models In addition to this, there are several other com-
for predicting rare events especially difficult. plexities: not all parts of a machine fail with the
First, due to the very fact that they are rare, the same frequency, the type of failure can vary, and
training data does not contain many examples more. Because sensor data is streamed in real
of these events. Given that such events are often time, it can have missing values or garbage
complex (for example, the sensor readings at the values as well. The frequency at which each
time of failure may not point to the cause of fail- sensor transmits data can also vary. Hence data
ure, since the failure could have been the result preparation and modeling techniques are needed
of a buildup of several other events already to help prepare the data and build predictive
reported as past sensor readings), it is often models to address the nature of the problem and
not easy to generalize from the few available capture the necessary information to effectively
examples. Second, depending on how predictive predict the failure or breakdown of the machine.
algorithms are configured and what underlying

Accurately predicting the probability of

occurrences of rare events like machine failures
requires robust algorithms.

2015 SAP SE or an SAP affiliate company. All rights reserved.

5/12
Using Predictive Maintenance to Approach Zero Downtime

Predictive Maintenance Use Cases

TRANSPORTATION For fleet owners and public transport depart-

The cost for maintaining an aging fleet of com- ments, it is important to have their vehicles
mercial, private, or military vehicles is high, but running all the time. Predictive analytics on
so is the cost of repairing relatively new vehicles real-time sensor data streams combined with
when they break down. There is a great benefit in fleet management systems will make it possible
both cases to reducing the probability of a failure to plan predictive maintenance for fleet vehicles
by performing preventive repairs and extending while they are on the road. This will help ensure
the useful life of the vehicle. Telemetric data that the vehicles are maintained at whatever
coupled with predictive techniques makes this location they are in, avoiding significant down-
possible today. Telemetric data from thousands time and financial losses due to breakdown.
of sensors captures the real-time status of the
various parts of a vehicle, which can be analyzed MANUFACTURING AND PRODUCTION
for actionable insights to improve the vehicles In manufacturing and production industries
safety and performance. For example: employing massive automatic machines, the
Vehicle manufacturers can use predictive ana- health of the machines has a significant impact
lytics on sensor data to find the key factors that on the quality of the products produced. In addi-
influence the fuel efficiency and performance of tion, any machine downtime can lead to a delay
their vehicles and thus improve the engineering in the production cycle and create a significant
design. bottleneck in the entire production flow, resulting
Vehicle manufacturers can use sensor data in in heavy costs. The profit margins of a company
combination with geolocation to understand can also be severely dented when machine fail-
the patterns of vehicle failure in different ures result in substantial downtime.
regions and then strategize what services to
focus on in each region, and what spares to Technological advances have made it possible to
stock up on to give their customers the best collect huge volumes of sensor data and evaluate it
service. Geolocation can be used to indicate to identify anomalies in the readings. For example:
to a connected car driver that there are likely Predictive modeling techniques can be
potential problems with the car and it should employed to predict the failure of a part far
be serviced at the nearest station. ahead in time and proactively replace the parts
Vehicle sellers can analyze sensor data to at the most convenient time in the production
predict the probability of a breakdown far ahead schedule to avoid any downtime.
in time and proactively reach out to their cus- Predictive modeling can be used to identify
tomers to perform maintenance keeping the the spare parts and the required tools that
vehicles running and customer satisfaction need to be stocked for minimizing downtime
high. in manufacturing.

2015 SAP SE or an SAP affiliate company. All rights reserved.

6/12
Using Predictive Maintenance to Approach Zero Downtime

During the phases of production, there are UTILITIES

often several paths that the product goes Power company operators and distributors face
through before it is defined as scrap. Predictive significant problems due to transmission losses,
modeling can help identify the sequence of meters that have been tampered with, and spikes
paths that result in a high probability of a part in demand. The distribution network is spread
being scrapped, and bring this to light far ahead across large regions, and the network, genera-
in the production cycle, resulting in significant tors, transformers, turbines, and smart meters
savings. are monitored for malfunctions. Any breakdown
Testing various mechanical parts during pro- could result in significant revenue loss for an
duction is critical to ensure that the machines entire industrial sector or government. Similarly,
built with them are of the best quality and have with the significant advances in instrumentation
optimal performance. For example, engines technology, oil fields, rigs, and refineries have
built for airplanes must be tested rigorously now become more digital. Data generated by
before they can be used. During production of thousands of sensors monitoring the functioning
these engines, sensors mounted into them help of various parts of the oilfield is streamed in real
monitor their hundreds of components and time to a centralized location. Predicting the pos-
capture the interactions of various engine sibility of a malfunction of any piece of equip-
parts. Predictive modeling can be employed to ment is critical so that corrective action can be
identify patterns that would result in an engine taken before a disaster occurs. Predictive model-
failure during the production phase itself, thus ing can be employed on data that is captured
allowing necessary changes to be made. in real time from the sensors connected to the
distribution network, for predictive maintenance.

In manufacturing, any machine downtime

can delay the production cycle and create a
significant bottleneck in the entire production
flow, resulting in heavy costs.

2015 SAP SE or an SAP affiliate company. All rights reserved.

7/12
Using Predictive Maintenance to Approach Zero Downtime

For example: DATA CENTERS AND CLOUDS

Predictive maintenance on assets such as gas Data volumes are growing rapidly, and organiza-
pipes and electrical cables used across the grid tions want to store all this data to mine informa-
can be an efficient and effective way to keep tion for any number of use cases in addition to
costs down, risks low, and availability high. predictive maintenance. This data is business
Predictive modeling can be used to identify the critical and is typically stored on an elastic infra-
probability of smart meters being tampered structure in data centers. Many organizations are
with and to plan preventive checks and mainte- also now moving to the cloud because it is easy
nance of those meters. to make new services available to their clients,
Sensor data from smart grids can be used for who benefit from the pay-per-use model. It is
optimizing the grid and for identifying potential essential to have the data centers and cloud
failures in the distribution network and schedul- infrastructure up and running all the time.
ing planned preventive maintenance to help Unstructured data in the form of logs containing
ensure zero downtime. the status of various services running on the
infrastructure, available in real time, can be
MEDICAL EQUIPMENT mined for predicting the failure of various ser-
The breakdown or malfunction of certain diag- vices and performing preventive maintenance to
nostic and lifesaving or life-support equipment in achieve zero downtime.
medical and pathology labs can be catastrophic.
Sensors within certain types of equipment are
used to record human body responses to diag-
nose critical illness so that appropriate care and
medication can be administered. It is therefore
critical to monitor this equipment in real time to
predict the probability of malfunction or break-
down and schedule preventive maintenance.

2015 SAP SE or an SAP affiliate company. All rights reserved.

8/12
Using Predictive Maintenance to Approach Zero Downtime

Enabling Predictive Maintenance

SAP PREDICTIVE ANALYTICS The data manager component in SAP Predictive

SAP Predictive Analytics software has capabili- Analytics enables the user to create ultrawide
ties that make it a product of choice for building analytical data sets for a time-stamped popula-
robust models for predicting rare events in the tion, and automatically extract thousands of
use cases just described. derived features that can be used for creating
robust models.
Predictive maintenance data is time series data
derived from the readings from thousands of If the event describes whether the machine has
sensors streamed at high velocity at real time. failed or not, then the automated classification
Sensor readings reflect the state of the machine component in SAP Predictive Analytics can be
performance at that time. So the time series data used to identify the chances of machine failure
must be transformed into a time-stamped popu- based on the ultrawide data set generated by the
lation, and new derived features must be defined data manager. On the other hand, if the event is
to find out what changes in the sensor readings recorded as a continuous number, then the auto-
over a period of time provide maximum informa- mated regression component in SAP Predictive
tion about the probability of failure of a machine. Analytics can be used to identify the relationship
One of the several ways of modeling data is of the sensor with the target. The classification
shown in Figure 1. Various statistical signatures and regression algorithms can identify the key
of each sensor are captured for a specific look- influencers that have the most information to
back period and associated with a segment of explain the behavior of the event that is being
data from the end of that period, to capture infor- modeled. They also identify the positive, negative,
mation of an event (such as machine failure) and neutral impact of the range of values of each
during that segment to use in building predictive sensor involved in the event, enabling the user to
models. Several such patterns can be generated have an in-depth understanding of the causes
at a step size equal to the forward segment size. leading to the modeled event.
The forward segment size defines the duration of
the validity of the forward prediction that is, if
the segment size is 10 minutes, then the model
built will be able to give the probability of an
event occurring in the next 10 minutes.

With SAP Predictive Analytics, the vast majority of

predictive maintenance use cases can be handled
automatically by a person with very little data
science expertise.

2015 SAP SE or an SAP affiliate company. All rights reserved.

9/12
Using Predictive Maintenance to Approach Zero Downtime

Figure 1: Data Preparation for Streams of Sensor Data

Raw sensor data stream

Processed sensor data
Time stamp Machine ID Sensor ID Sensor value 1
18/01/2009 10:32:46 1020010 Sensor 1 0.924387
Machine ID: 1020010
19/01/2009 10:32:46 542 Sensor 4 1.2235
20/01/2009 10:32:46 1020010 Sensor 3 3.397467 Time stamp Sensor 1 Sensor 2 Sensor 3 ... Sensor n
21/01/2009 10:32:46 542 123.876 18/01/2009 10:32:46 0.924387 3.600801 3.3997467 3.569223
22/01/2009 10:32:46 1020010 Sensor n 3.569223 18/01/2009 10:32:47 0.280205 3.920714 1.524253 3.950853
23/01/2009 10:32:46 6482 Sensor 99 12887 18/01/2009 10:32:48 0.466136 1.23772 1.599199 2.662753
24/01/2009 10:32:46 1020010 Sensor 2 3.600801 ... ... ... ... ... ...
25/01/2009 10:32:46 . . . ... ... 18/01/2009 10:32:50 0.44493 3.652581 3.137181 0.80788

Billions of rows low data density

2 Data transformation
Look-back period Step size

Normal
event
Time series
Abnormal
data stream event

3 Ultrawide data sets

Time stamp Machine ID Type Num. Sensors Max_Sensor 1 Min_Sensor 1

Millions of rows
20/01/2009 10:30
thousands of columns 20/01/2009 11:00
high information density ... ... ... ... ... ...

2015 SAP SE or an SAP affiliate company. All rights reserved.

10/12
Using Predictive Maintenance to Approach Zero Downtime

The automated time series in conjunction with an The automated clustering component enables
automated regression algorithm helps in prob- the user to find robust numbers of clusters in the
lems related to forecasting, such as the remain- data and can be used for problems like grouping
ing time before the machine or its parts break machines behaving in a similar fashion and sub-
down in other words, the time to next mainte- sequent anomalies in the sensor data.
nance. This is shown in Figure 2.

Figure 2: Time to Next Maintenance

On/off

0
Time
Accumulated operating hours

Next service

Now

Predicted
date
0

2015 SAP SE or an SAP affiliate company. All rights reserved.

11/12
Using Predictive Maintenance to Approach Zero Downtime

All the automated components build several The models can be easily integrated into busi-
models internally and choose the most robust ness processes and can be scheduled for auto-
models based on their predictive power and matic retraining if there is significant deviation in
predictive confidence. Predictive confidence cor- the data or in the models performance. The vast
responds to the proportion of information con- majority of predictive maintenance use cases can
tained in the target variable that the explanatory be handled with an automated approach, and the
variables are able to explain in other words, the software can be easily used by a person with very
ability of the model to explain the target. Predic- little or no data science expertise. For the most
tive power is the capacity of the model to achieve complex problems, the expert analytics option in
the same performance when it is applied to a SAP Predictive Analytics provides a workbench
new data set exhibiting the same characteristics where the automated algorithms can be com-
as the training data set. This is the generalization bined with other prebuilt algorithms and with
ability or robustness of the model. custom algorithms written using the R program-
ming language.

LEARN MORE
SAP Predictive Analytics software is a powerful solution
designed to address several predictive modeling use cases
involving high-velocity data streams that capture machine
behavior in real time. To find out more about how this soft-
ware can help your company use predictive maintenance to
achieve zero downtime, please contact your SAP representa-
tive or visit www.sap.com/predictive.

Studio SAP | 39044enUS (15/07) 2015 SAP SE or an SAP affiliate company. All rights reserved.

12/12
 2015 SAP SE or an SAP affiliate company. All rights reserved.

No part of this publication may be reproduced or transmitted in any form or for any purpose without the
express permission of SAP SE or an SAP affiliate company.

SAP and other SAP products and services mentioned herein as well as their respective logos are
trademarks or registered trademarks of SAP SE (or an SAP affiliate company) in Germany and other
countries. Please see http://www.sap.com/corporate-en/legal/copyright/index.epx#trademark for
additional trademark information and notices. Some software products marketed by SAP SE and its
distributors contain proprietary software components of other software vendors.

National product specifications may vary.

These materials are provided by SAP SE or an SAP affiliate company for informational purposes only,
without representation or warranty of any kind, and SAP SE or its affiliated companies shall not be liable for
errors or omissions with respect to the materials. The only warranties for SAP SE or SAP affiliate company
products and services are those that are set forth in the express warranty statements accompanying such
products and services, if any. Nothing herein should be construed as constituting an additional warranty.

In particular, SAP SE or its affiliated companies have no obligation to pursue any course of business
outlined in this document or any related presentation, or to develop or release any functionality mentioned
therein. This document, or any related presentation, and SAP SEs or its affiliated companies strategy
and possible future developments, products, and/or platform directions and functionality are all subject
to change and may be changed by SAP SE or its affiliated companies at any time for any reason without
notice. The information in this document is not a commitment, promise, or legal obligation to deliver
any material, code, or functionality. All forward-looking statements are subject to various risks and
uncertainties that could cause actual results to differ materially from expectations. Readers are cautioned
not to place undue reliance on these forward-looking statements, which speak only as of their dates, and
they should not be relied upon in making purchasing decisions.