Physics-Informed Neural Networks with Mittag-Leffler Kernel for Gear Anomaly Prediction

This paper introduces a hybrid deep learning framework based on physical information for anomaly prediction in gear systems. The method integrates a Mittag-Leffler kernel convolution layer to simulate fractional order physical memory effects, which reflects the viscoelastic and non-local characteristics of gear dynamics. Long Short Term Memory (LSTM) networks and Transformer based modules are also employed to extract sequences and long range temporal dependencies. Additionally, a Physical-Information Neural Network (PINN) is embedded in the architecture, enabling the network to effectively generalize in simulated datasets that follow physical laws, and residual connections are used to enhance training stability. Experiments on physical based simulation datasets demonstrate that this model is highly effective in early anomaly prediction, which is beneficial for predictive maintenance and fault prevention of gear transmission systems.