Predictive Maintenance

Already using Condition-based Maintenance approaches for your devices, you utilize your IoT sensor data in order to measure performance analysis and predict the machines health status. Sensor-based monitoring can be applied to different aspects of data like vibration, temperature, voltage, radiation and mechanical wear data. However, conventional methods are struggling with situations of the following kind:

• Weak Labelling
  • Failures are not attributed to a single causal component.
  • Failures are not confirmed by experts (false positives in the target variable).
  • Machines are not repaired even though they are faulty (false negatives in the target variable).
• Low failure counts:
  • Expensive, rarely changed components make it hard to learn these faults.
• Complex dependencies of measurements:
  • Dependencies are often mathematically unknown and thus can’t be used for rigid models; instead they have to be approximated/learned via machine learning.

Our approach learns the expected dependencies of this network of aspects of a machine and leverages deviations from these baseline dependencies in order to predict the health status of the machine. We can utilize fine details by an intricate vibrational analysis as well as large spreads of these deviations. In this way we can observe the control units behaviour and draw conclusions about its reaction to specific component failures predicting which component is malfunctioning in a machine by its current control unit behaviour. This allows you to leverage further aspects of your IoT data for additional insights in your machines health status and each components longevity. Since our approach is semi-supervised and thus is not needing a failure diagnosis for its data preprocessing it can handle small numbers of failures. Furthermore by using suited algorithms and filtering based on the weak label we are able to handle cases in which the label contains false negatives and false positives.