Feature-Based Fault Detection Approaches

With increasing complexity of systems it is becoming more and more time consuming and difficult to achieve reliable fault detection strategies. Using model-based methods requires detailed knowledge about the systems behavior and seems in some cases successful in theory but un-applicable in real-time due to high computation requirements. In this contribution, an idea and algorithm for feature-based fault detection approach is proposed. The main idea of this approach is to detect and identify faults in a complex system without any kind of modeling. By extracting features from relevant sensor signals, yielded from hardware-in-the-loop simulations, and combining them in a matrix, it is possible for a human operator to denote subsets of the matrix as fault-free and faulty areas. An advantage is the ability to set individual thresholds for the subsets, giving a more robustness towards false alarms and a possibility to denote individual subsets to relevant faults. From this, it will be shown that identification of faults is possible. In order to achieve a fault detection and identification ability, it is necessary to implement the faults of interest in a test rig and conduct hardware-in-the-loop simulations. The raw data from fault-free and faulty simulations are used in the training of the matrix and the algorithm detects and identifies the faults in a robust way. The results are compared to a classical fault detection method that uses fixed thresholds. It will be shown how a sensor bias fault and a pressure relief valve fault are detected and identified