Fault Diagnosis of Controllable Pitch Propeller as Few-Shot Classification with Mechanism Simulation Data Augmentation

The Controllable pitch propeller(CPP) can control the thrust direction and power through the adjustment of blade pitch using a hydraulic system, significantly improving the maneuverability of the vessel. Given the harsh working environment of the on-ship CPP hydraulic system, it is prone to failure and affects pitch control performance. To address the challenges of limited real-world fault condition data, we utilize a few-shot learning approach and enhance the model's sensitivity to target data distribution by introducing a simulation data augmentation phase during training. A mechanical model of the CPP hydraulic system is developed, and monitoring data under fault conditions of key components is simulated. After training with open-source time-series classification tasks, the augmentation phase is used to obtain prior knowledge of the equipment, and the model is eventually adapted to the recorded data task. The recorded data is derived from an experimental setup comprising the primary components of the CPP, and it is collected from multiple sensors under both normal and fault conditions. The target fault diagnosis task is treated as a time-series classification problem, and a network architecture based on ResNet with Triplet loss is employed.