Recognizing Anomalies

Discovering outliers

Types of anomalies

Anomalies are individual measurement points or groups of measurement points that represent irregularities in data. For example, an anomaly can be an unexpectedly high torque measurement in the initial phase of a production process. Temporary irregularities in the development of measured variables over time, such as those found in an ECG image in cases of cardiac arrhythmia, are also anomalies. In general, whether a single measurement or the chronological sequence of a measured variable is anomalous depends on the context or circumstance. For instance, a high torque measurement in the initial phase of production could be anomalous but, in the end phase, it would be absolutely normal.          

Methods to recognize anomalies

There are many different ways to recognize anomalies, which vary depending on the type of data, the type of anomaly, and the domain knowledge that is available. One possibility is to use cluster analysis methods, which classify data points as outliers that cannot be assigned to any identified cluster. Other approaches, especially those used to analyze time series data, use recurrent neural networks, which predict the current measurements based on past measurements. If the predicted value and the actual value differ too much, the measurement is classified as an outlier.           

Anomaly detection of sensor data in real time

Identifying and eliminating outliers during operation is extremely important, especially for real-time analysis systems. For example, this prevents measuring errors being reproduced in subsequent analysis steps and falsifying results. As opposed to detection methods based on classic cluster analysis processes, the solution shown here allows outliers to be identified and eliminated in real time.     

The illustration shows a snapshot of outlier identification from sensor data inflowing in real time. Measurements outside the tolerance range (light blue) are identified as outliers (red) and removed. The moving median is inserted in place of the outlier.

Your benefit from the recognizing anomalies

Recognize errors
in good time




Any questions? We are happy to help.

Katana, the data analytics segment of the USU Group, has extensive experience in the application of machine learning methods in the industrial sector. Take advantage of our knowledge and our solutions for building your data-driven business models to reduce your costs and improve your quality and value added.