If sensors suffer temporary faults when taking a measurement (see blue curve in the defective section), this can often render the entire measurement useless. Because taking measurements can be a very time-consuming and costly process, these series of measurements should not be discarded if possible.

With the help of solutions developed by us, defective measurement intervals can be corrected (see blue and green curve in the defective section) and measurements can be adjusted for drift (see red curve).

Identifying and eliminating outliers during operation is extremely important, especially for real-time analysis systems. This helps to prevent errors caused by outliers being reproduced in subsequent processing steps and thus falsifying analysis results.

Unlike the classic cluster analysis method, our solution allows outliers in sensor data to be both identified and eliminated during operation.

Anomalies are irregularities in the measurement data from machines and therefore play a key role in identifying faults.

With the help of the anomaly recognition system developed by us, historical data are used to automatically identify various machine operating phases and to calculate phase-specific tolerance thresholds. The sensitivity of the anomaly recognition system can be set by the user on an individual basis. As a result, even slight deviations from normal machine behavior can be identified in real time while the machine is operating.

In addition to real-time anomaly recognition in sensor data, a long-term prediction of the development of relevant key variables over time plays a central role as well. Key variables could be, for example, a machine’s performance or oil quality. If the development of these key variables over time can be predicted with sufficient accuracy, spare parts can be ordered on time and service technicians can be scheduled for maintenance. This helps to prevent costly downtimes.

Prediction approaches based on a linear extrapolation of measured variables can lead to highly inaccurate predictions. The approaches developed by us are based on machine learning procedures. They offer improved predictions with additional information regarding the uncertainty of the prediction. See illustration.

Both complex machines and IT systems describe their status by outputting discrete events in the form of log data. These generally contain important information regarding the status of the systems and developments over time. Thanks to self-learning procedures, our solutions are able to structure log data and identify the most likely causes of the occurrences of sudden events. As a result, we reveal any relevant links.