Detecting unusual conditions under varying loads
IAP developed a decision support system for condition-based maintenance with DB Cargo’s predictive maintenance department. Operating and sensor data was to provide indications of possible component faults, helping maintenance staff investigate unusual conditions early.
Sensor readings change with load even during normal operation. Sensors also often measure the behaviour of the whole machine while the underlying cause lies in an individual component. The approach therefore examines how measurements relate to each other and when those relationships change.
Working with rare faults and uncertain fault labels
Training data usually contains far more examples of normal operation than confirmed faults. Failure examples are particularly scarce for components that are rarely replaced. The link between a repair record and the actual cause of a fault can also be uncertain.
IAP investigated maintenance experts’ hypotheses using historical measurement and maintenance data. The investigation examined whether the conditions of interest could be distinguished in the data and how informative the available fault labels were.
Normal operating behaviour as the reference
Using operating conditions assessed as healthy, the model learns the expected relationships between load and the machine’s response. These relationships can be complex and are learned from the measurements. The method subsequently flags deviations from that expected behaviour.
The approach treats longer-term changes and oscillations separately. This allows both larger shifts in operating behaviour and finer dynamic anomalies to be investigated. The central question is whether the current load explains a deviation or whether it indicates a change in condition.
The semi-supervised approach uses the more abundant normal-operation data and supplements it with the available fault indications. This makes investigation possible even with few failure examples. Evaluation takes account of the possibility that existing labels contain false alarms or missed events.
Information for maintenance decisions
The method provides indications of possible component faults and visualises the conditions studied for further assessment by maintenance staff.
The visualisation shows the operating conditions in which anomalies occur and which changes merit closer investigation. This supports the selection of further checks and preparation of maintenance decisions. In his complete original statement, Thomas Pohl confirms the insights gained and the automated prediction function.
