Supply Chain Management Blogs by SAP
Expand your SAP SCM knowledge and stay informed about supply chain management technology and solutions with blog posts by SAP. Follow and stay connected.
Showing results for 
Search instead for 
Did you mean: 
Active Participant
Introductory note: just in case you are wondering about the product mentioned in the title; the product previously known as SAP Predictive Maintenance and Service (PdMS) as of the 2011 release now includes the simulation capabilities of SAP Predictive Engineering Insights enabled by Ansys, and was renamed to SAP Predictive Asset Insights (PAI).


A brief recap of machine learning in Predictive Asset Insights

Over the years we have assembled a rich set of machine learning capabilities in PdMS and now PAI. This includes:

  • The set of out-of-the-box algorithms for supervised and unsupervised training and scoring, like Logistic Regression or Principal Component Analysis

  • The “super-algorithms” which on their own churn through several of the individual algorithms, for Automatic Anomaly Detection and Automatic Failure Prediction

  • Failure Mode Analytics, which allows you to do a text analysis of the long text of a notification, and the description of a failure mode, and have the system propose which failure mode should be assigned to which notification, and display occurences

  • Leading Indicator Analytics, which allows you to determine which of the sensor data streams (indicators), and for which specific conditions (e.g. “Inlet temperature > 125 degrees Celsius”), may have contributed to a breakdown notification. As a bonus such conditions can be turned into a new rule in the PAI rules engine.


New in PAI 2011: Failure Curve Analytics

As a new machine learning feature Failure Curve Analytics (FCA) rounds out the ensemble of existing algorithms. It is based on a traditional Weibull model. This is very interesting to customers who have no, or incomplete, or inconsistent sensor data, as FCA only needs breakdown notifications to calculate a Probability of Failure (PoF) curve, and can achieve good results even with few notifications.

Using FCA works in in four steps.


Step 1: create a new FCA model configuration and set general parameters

In PAI 2011 you will find in the Machine Learning Engine group a new application for “Failure Curve Analytics Model Configuration”

Open it to see a list of existing model configurations, and create a new one. Here you can enter, apart from its name and description, a few key parameters:

  • Age for Conditional Probability of Failure – FCA will always calculate a PoF curve for the fleet of selected equipment, but entering a number here will also calculate what the PoF is for the equipment having survived thus far, and surviving until that target age

  • First Installation Data is Maintained and Reliable – sometimes customers have incorrect or no information when an equipment was installed, i.e. when its initial age was zero. If this is the case FCA will allow to use the end of the first breakdown as the “birth” of the new equipment.

  • Equipment is Repairable – FCA assumes that after a repair an equipment is in an as-new state; if that is not the case the equipment is assumed to be scrapped.

Selecting these parameters influences what FCA calculates, and how it calculates it.


Step 2: selecting input business data

FCA works with three sets of input data:

  • A fleet of equipment – FCA allows the user to hand-pick which equipment which have been exposed to the same operating conditions shall be analyzed together. For this FCA offers a filter dialog with a few standard equipment attributes, and any custom attributes the customer may have added to the equipment.A fleet will be stored in a PAI “Fleet Group” (there is a separate application to view and edit them), and existing fleet groups can be reused for multiple FCA model configurations.

  • A set of failure modes – failure curves are always calculated per failure mode. The system will collect any failure modes assigned to the fleet of equipment, and the user can select all or some of them.

  • A date range of breakdown notifications – FCA will collect all breakdown notifications for the selected failure modes, for one or multiple date ranges

Step 3: train and score the model

In FCA we have introduced a handy new feature for machine learning applications: a single button which will train and score the FCA model configuration in one fall swoop!



If you want to or need to you can now watch the information / warning / error logs of the training and scoring run. What you will hope to see is a log without errors:



The training will calculate the parameters for a Weibull model, for the whole fleet of equipment, and the scoring will compute all the other output parameters, such as the age of the equipment, the time-to-failure, etc.


Step 4: view the output data in an equipment chart

When the training and scoring is done the user can select one of the equipment to navigate to its “object page” in the equipment application. There we have added a new entry in the Analytics section:



In that section the user will see a PoF curve. It is specific to (a) a model configuration (an equipment could have been used in multiple models, e.g. one for “All pumps older than 3 years” and one for “All pumps in location ABC from manufacturer XYZ”), and (b) a failure mode. The user can select from either.


The chart shows the current age of the equipment and the PoF curve, including a confidence interval.



The use can select a point on the chart for its specific age/PoF combination.



Selecting the point on the curve for the current equipment age also shows the calculated time-to-failure and that date.



Selecting a different model configuration or failure mode will show those results.



Using the information

Using the chart data they user can estimate how quickly the equipment may be in a critical situation.


More information

SAP Help is a good source for more information:


What else?

We have a solid set of additional features in mind (e.g. to write the results of the model into an indicator), but as this is fresh new functionality, we also plan to let customers play with it for a release, to see how well it works against their data. For some customers using Weibull analyses is common practice, but for other customers this is new ground.


What do you desire?

How about your company? Have you used Weibull models before, and what worked for you, and what not? What features are you missing? In which business processes did you use it? Would you want to use Weibull analyses in SAP applications?

I’d love to hear from you!