Anomaly detection

The importance of recognizing
ANOMALIES IN COMPANY DATA

Whether it's human behavior or measurements from IoT devices, in the field of data science, there can always be deviations from the expected results.
These unexpected deviations are called anomalies and can indicate potential problems, the identification of which can be of great importance in numerous business areas:

  • Recognizing attempts at fraud in orders
  • Monitoring IoT sensor data and log files
  • Ensuring compliance with purchasing processes

What is the core issue?
ANOMALIES REQUIRE SWIFT ACTION

The caused Costs of an anomaly naturally depend on the business area in which they occur. However, another significant factor is the time it takes for them to be discovered:
  • Think of a bug in software as an anomaly.
  • The graph illustrates the relationship between the discovery of a bug in a software project and the cost of fixing it.
With our showcase, we want to present you with a solution approach, to minimize the time between the occurrence and discovery of an anomaly.
anomaly detection_costs_errors
National Institute of Standards and Technology (NIST) | isr.de
Anomaly detection refers to the problem of finding patterns in data—
—that do not conform to expected behavior.
Chandola et al. (2009):
Anomaly detection: A survey

OUR SHOWCASE

Anomalies in machine log data

The Getinge, Maquet, and Atrium logos are trademarks of Getinge.

In our showcase, we use a concrete example to demonstrate how anomalies can be detected using artificial intelligenceand deep learning models.

Learn more now

The detection of data anomalies in a business context is complex.

Anomaly Detection Data Management
Many disciplines of data management are involved in the detection of anomalies | isr.de

Different forms &
TYPES OF ANOMALIES

ISR Logo
Punctual anomaly

A data record is considered abnormal when compared with the rest of the data.

ISR Logo
Contextual anomaly

A data record is considered abnormal in a specific context.

ISR Logo
Collective anomaly

Several data sets are considered abnormal when compared with the rest of the data.

Not all roads lead to Rome
FROM DATA TO MODEL ARCHITECTURE

Each use case requires an individual approach to the design of the machine learning model, depending on the boundary conditions and the type of anomalies to be expected. Key questions that we will clarify in consultation with you are:
  • Are the causal relationships in the data known?
  • Is there training data with already detected anomalies?
  • What types of anomalies can be expected?

The most important insights from our showcase "
" SUMMARIZED FOR YOU

In our showcase, we demonstrate how anomalies can be identified using modern deep learning models . Our solution has proven itself in practice and has been able to detect even anomalies that are difficult to detect. The ISR solution approach can be used immediately and can also be adapted to your requirements with little effort.

The algorithm achieves adetection rate of 90% in identifying artificial anomalies in the test data set.

Real anomalies could also be detected, but their rarity precludes reliable evaluation.

Combining different machine learning methods can further increase accuracy.

We see the biggest advantage in the fact that no labeled anomalies are required in advance for reliable detection. This saves a large amount of effort compared to other approaches.

The ISR balance sheet to date

59
Measured values for analysis
5
tested deep learning algorithms
59
Tested hyperparameter combinations
6
Different types of anomalies
The increasing volume of data in companies and the analysis of this data in real time make the use of machine learning algorithms increasingly attractive.
In the area of anomaly detection in particular, these algorithms can provide insights into interesting business transactions and reveal security gaps for Getinge.
Arnold Fritz
Senior Director IT Technologies | getinge

Take advantage of our expertise

ANOMALY DETECTION AS PART OF YOUR DATA SCIENCE PORTFOLIO

Data analysis & problem identification

Uncover problematic anomalies before they can cause damage to your business.

Pie chart with three equal sections in different shades of blue. Two of the three sections are not visible.

model training

Benefit from our experience—from the theoretical concept to the finished implementation.

Pie chart with three equal sections in different shades of blue. One of the three sections is not visible.

end-to-end integration

Don't start from scratch – with our showcase, we have a framework that can be adapted to your use case.

Pie chart with three equal areas in different shades of blue
Pie chart with three equal sections in different shades of blue. Two of the three sections are not visible.
Pie chart with three equal sections in different shades of blue. One of the three sections is not visible.
Pie chart with three equal areas in different shades of blue
01
Data analysis & problem identification

Uncover problematic anomalies before they can cause damage to your business.

02
model training

Benefit from our experience—from the theoretical concept to the finished implementation.

03
end-to-end integration

Don't start from scratch – with our showcase, we have a framework that can be adapted to your use case.

Our data science toolkit in the area of
DATA ANOMALIES

IBM Watson Studio
Data science & AI—simple and everywhere with IBM
Learn More

Other tools we use

TensorFlow ↗

Keras ↗

You might also be interested in
Our Showcase

Anomalies in machine log data
Download our showcase now
Download here
split image with close-up of one eye
Portrait Photo Silvio

Silvio Bergmann
Senior Manager
Analytics & Insights
silvio.bergmann@isr.de
+49(0)151 422 05 418

Xing