SAP Predictive Analytics Addressing Predictive Maintenance with SAP Predictive Analytics Table of Contents 2 Extract Predictive Insights from the Internet of Things 2 Design and Visualize Predictive Models Automatically 4 Why SAP Predictive Analytics? Check Out These High-Value Use Cases 4 Scenario 1: Identify Which Sensor Values Explain an Incident 7 Scenario 2: Identify Faulty Sensors and Incorrect Readings 10 Scenario 3: Identify Machines That Act in a Nonstandard Way 11 Conclusion
Advances in technology have enabled a connected world called the Internet of Things (IoT). Predictive maintenance is one of the prime use cases for the IoT. With smart sensors, you can now remotely monitor equipment 24x7 and assess the risk of failure. But as you deploy more and more sensors, huge amounts of data become available. SAP Predictive Analytics software lets you extract insight from this data by building robust models that help you understand changes in machine behavior in real time. EXTRACT PREDICTIVE INSIGHTS FROM THE INTERNET OF THINGS Now that processors, electronic components, sensors, and monitoring devices are significantly smaller and more energy efficient, the Internet of Things is truly ingrained in our lives (see Figure 1). It s at work across multiple vertical market segments, including automotive, consumer electronics, healthcare, manufacturing, transportation, and utilities, just to name a few. What s more, the decreased cost of sensors makes it economically viable to integrate large numbers of them into a physical grid for collecting machine-related performance data. This data has enormous potential for predictive maintenance but only if you can use it to learn or discover useful information and actionable insights in real time for maximum impact. All too often, however, the sheer volume of collected data becomes an issue. Today thousands of sensors can send data each millisecond for the purpose of monitoring a single machine. Moreover, many organizations add as many sensors as possible without proof of their usefulness either to compensate for possible sensor issues or to monitor every part of the machine. As a result, vast quantities of information accumulate in real time, which makes the development of a predictive model a challenging task. The ability to react in real time is reduced due to the amount of data that needs to be processed. DESIGN AND VISUALIZE PREDICTIVE MODELS AUTOMATICALLY SAP Predictive Analytics provides a solution for your predictive maintenance projects. In addition to optimizing sensor performance, it decreases the number of sensors needed to create a predictive model and monitor a machine over the long run. The software automatically builds various models that point out the most important readings, indicate which anomalies should be controlled, reveal how to avoid redundant sensors, and identify how to control the machine in real time. 2 / 11
Figure 1: The Internet of Things Landscape Connected home Responsive supply chain Connected cities Predictive maintenance Connected buildings Connected cars Connected logistics 3 / 11
Why SAP Predictive Analytics? Check Out These High-Value Use Cases The following scenarios show how SAP Predictive Analytics software can help optimize predictive maintenance projects, reduce cost, and focus attention on the most important issues. SCENARIO 1: IDENTIFY WHICH SENSOR VALUES EXPLAIN AN INCIDENT How can you reduce the number of sensors that you re monitoring without jeopardizing accuracy when predicting issues? In this scenario, thousands of sensors are connected to a machine. The machine is subject to vibrations caused by the speed of the motors, the thrust of the pistons, the effect of electromechanical switches, and the movements of mechanical parts. Sudden changes in the normal vibrations are considered abnormal. The aim is to find which parts of the machine you need to tune in order to reduce the overall machine vibrations. SAP Predictive Analytics uses streams of sensor data from several comparable machines under different operating conditions for modeling. The software enriches the raw sensor data taken at different time lags with derived attributes using statistical signatures such as sensor maximum readings, minimum readings, and deviations in different time windows. machine, you can use the rate of change in the frequencies measured for vibration as the target. A threshold is kept to define the target. The rate of change in vibration frequency greater than the threshold is set as Class 1, which is the class of interest (the target) for this scenario; the other records below the threshold are set as Class 0. In this scenario, let s say that approximately 3% of the records belong to the observed target (Class 1) in the data set (see Figure 2). The automated analytics classification algorithm that is built into SAP Predictive Analytics is robust enough to model such scenarios where the observed target occurrences are very small. Using this algorithm, you can find the sensors that best describe the sudden increase or decrease in the vibration of the machine. Figure 2: Sudden Increase Versus Decrease of Vibration As the overall intent is to find the sudden increase or decrease in the vibration of the 4 / 11
The classification model provides a list of variables and their contributions toward identifying the target (that is, the sudden increase and decrease of the vibrations). The most influencing variables are called Key Influencers. A preview of the report generated by the automated analytics component of SAP Predictive Analytics is shown in Figure 3. As you can see, this report identifies the most important sensors to monitor. Figure 3: Smart Variable Contribution Report 5 / 11
As shown in Figure 4, you can also use the automated analytics classification algorithm to find the range of values of the sensors that are most or least influential for the behavior being modeled. In other words, this report identifies which sensor values are most likely to indicate a sudden change of vibration. Note: Predictive algorithms use historical data for modeling. Predicting machine failures based on historical information might be a significant challenge. This is because the absolute number of failures is significantly low given the advances in technology and planned maintenance of machines. For example, 20 records of sensor readings for machine failure compared to 100,000 records for a normally functioning machine are insufficient to build a good quality model. But as failure is very costly, it is useful to identify abnormalities in the sensor readings and find early signals in the data to prevent the occurrences of such extremely rare events. Figure 4: Variables Contribution Report Less contributive Variables contribution Most contributive [-235.406 ; -20] [-61.563 ; -4.531] [-2816 ; -1858] [-61.563 ; -4.531] [-31.313 ; -25.563] [-61.563 ; -4.531] [-15.969 ; 2.495[ [-30.859 ; 3.84] [-34 ; 36] [-2698 ; -1910] S454 S531 S319 S530 S248 S529 S544 S483 S541 S488 ]237.406 ; 269.438[ ]55.156 ; 85.188] ]3457 ; 4321] ]55.156 ; 85.188] ]238.438 ; 286] ]55.156 ; 85.188] ]989.567 ; 1433.2] ]1978 ; 2869.14] [1942 ; 2869] ]1844 ; 7596] 6 / 11
SCENARIO 2: IDENTIFY FAULTY SENSORS AND INCORRECT READINGS How can you identify malfunctioning sensors that are providing inaccurate readings? Instead of monitoring all the readings, you can automatically create a model showing the influence of other sensors on the one being monitored and raising an alert when the sensor shows discrepancies from the expected readings. The prediction of the exact time of failure is a challenge because a very small number of histor- ical observations show failures of a machine or machine parts. You can instead find relationships in the sensor readings of a machine that is working in perfect condition. All the parts of a machine work in a highly synchronized manner, so it is possible to find robust relationships between sensors either in real time or at specific time lags. With its built-in automated analytics regression algorithm, SAP Predictive Analytics can find the relationship between the sensors. Figure 5 shows a snapshot of the visualization of readings from a few sensors. Figure 5: Sensor Reading Visualization 7 / 11
Figure 6: Regression Model for Sensor 1 Sensor 1 408.56 407.09 394.70 395.93 392.47 396.51 Sensor 2 Sensor 3 Sensor 4 Sensor 5 168139.65 13.69 63745882.05 65.96 144779.86 14.09 65950294.65 60.52 173894.12 14.17 61224929.10 54.67 158865.39 14.08 64638241.68 63.33 154304.08 13.67 62260010.21 56.90 165413.39 13.94 62510230.76 58.35 Regression model You can also use regression models to find relationships between the readings of a particular sensor and the readings of all the other sensors. With SAP Predictive Analytics, you can build a large number of regression models: one for each of the sensors (see Figure 6). The automated analytics regression algorithm automatically finds the leak variables (that is, the individual sensors that can completely explain the target sensor for which the regression model is being built). In addition, the algorithm automatically removes highly correlated sensors in order to find robust regression models. It also identifies the most influencing variables (that is, the most important sensors that best explain the target sensor used for modeling). This makes it possible to significantly reduce the number of sensors that you need to evaluate in order to find anomalies in the sensor data. In real time, you can use the regression models built using the important sensors to predict the expected readings of a sensor and compare them with the actual readings. A significant deviation between the actual and predicted values of the target sensor indicates some abnormal behavior of the sensor reading in real time (see Figure 7). Figure 7: Anomaly Detection Based on Regression Models Diff = sensor actual sensor predicted Probability model Anomaly detection 8 / 11
In this scenario, the number of models created depends on the number of sensors streaming data in real time from the machine. It is very cumbersome for the data scientist to build thousands of models manually. Automated analytics eliminates this complexity and creates massive automation for such use cases. The model manager of SAP Predictive Analytics helps you deploy, manage, and monitor these models in a production environment (see Figure 8). Figure 8: Model Manager 9 / 11
SCENARIO 3: IDENTIFY MACHINES THAT ACT IN A NONSTANDARD WAY This scenario shows how you can automatically define a group of machines that have normal behavior and identify those that are behaving differently (that is, they have sensor readings that deviate from the usual ones). By analyzing those outlier machines, you can quickly identify possible defect areas. You may even find that some nonstandard behaviors are actually providing better functionality that you want to deploy on all your normally behaving machines. In general, the sensor readings from machines that are behaving normally follow a similar pattern. As a result, it is possible to group the machines together based on the sensor readings. The automated analytics clustering algorithm that s built into SAP Predictive Analytics lets you specify the minimum and maximum number of clusters in which data should be categorized. The algorithm automatically finds the best number of clusters from among the range of clusters specified. After running the algorithm, you can study each cluster to understand the behavior of the machines belonging to it and label the cluster according to the condition of its machines. The goal is to have an overall understanding of the nature of the clusters. Once you have determined a clustering model, you can assign the machines in real time to one of the nearest clusters based on the new sensor readings. You can identify an abnormal behavior of a machine when it is assigned to a cluster representing erroneous behaviors. Alternatively, you can recluster machines again and again at predefined time intervals based on new sensor readings. This enables you to find machines that are outliers or unassigned to any clusters due to abnormal patterns captured by the sensor readings as shown in Figure 9. Figure 9: Outlier Detection Based on Clustering Outlier machines 10 / 11
CONCLUSION Now you can easily build models covering several predictive maintenance scenarios using the automated analytics module of SAP Predictive Analytics. Along with the various automated algorithms shown here, SAP Predictive Analytics contains additional modules that extend its use to address other predictive maintenance needs. These modules include: Data manager To automate the definition of a data set enriched with thousands of derived variables (such as the past values of sensor data) Expert analytics To let users define their own algorithms or reuse existing libraries for the SAP HANA platform or open source R environment Model manager To automatically maximize model quality by comparing predictions and new actual data; whenever the quality goes below a given threshold, the software alerts and retrains the model You can find more information about SAP Predictive Analytics at www.sap.com/predictive. Studio SAP 39890enUS (15/09) 11 / 11
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 SE s 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.