Sample generation method for marine diesel engines based on FEM simulation and DCGAN

The healthy and stable operation of the ship's power system is the foundation for the normal navigation of a ship. Data-driven ship power system condition monitoring is currently one of the research directions, but such methods often require a large amount of labeled data support. How to obtain a sufficient number of fault samples is the first problem to be solved for such methods. Therefore, a new fault sample generation scheme is proposed, which first uses the finite element method (FEM) to generate vibration data of marine diesel engines in different fault states, and uses deep convolutional generative adversarial network (DCGAN) to narrow the domain difference between simulation data and measured data, while retaining the fault characteristics of the simulation data, thereby generating synthetic fault data that is closer to the real fault state. The iteration number is determined through the comparison of time-domain, frequency-domain, loss function changes, and fault type identification results of synthetic data, measured data, and simulation data. The quality of the synthetic data is judged, and ultimately, a high-quality data sample for model training is generated.