A context-aware methodology for fault detection and severity classification under variable operating conditions

Rotating machinery operates under variable conditions making monitoring particularly challenging, as often it is not clear if a change in the vibration signature is due to a fault or due to the change of operating modes. Therefore, the introduction of context in condition monitoring algorithms has been proposed in order to provide the necessary background for the correct interpretation of data and informed decision making. Context allows algorithms to tailor their analysis to the unique characteristics of the machine operation and its environment. In this paper, the problem of fault detection and fault severity classification in rotating machinery under variable and unseen working conditions is investigated. Existing state-of-the-art methods are often based on deep multimodal data fusion, combining contextual information (i.e. operating mode) with raw sensor data, to account for the additional information encoded in the operating states. However, this fusion increases task complexity, since models must recognize fault signatures across multiple operating modes in which machine behaviour can vary substantially. To address this challenge, we propose a novel hypernetwork based methodology. Our hypernetwork takes the current operating mode as input and generates the weights of a specialized prediction network tailored to that mode. This design enables efficient adaptation to -and knowledge sharing across- different operating conditions while maintaining a simpler per mode prediction model, thereby reducing overall task complexity. The methodology is tested and evaluated on the PHM2023 gearbox dataset and the results show that it achieves generalization and accuracy improvement in the prediction of fault severity under unseen operating conditions, outperforming state-of-the-art techniques.