Physics Guided Generative Learning for Domain Adaptable Data Synthesis : Progressive Fault Synthesization for Predictive Machine Maintenance

In this work, we perform Physics guided data synthesis. Proposed method uses seed data from the observed source state / domain for data generation of unobserved target state / domain. Our method adapts the variation of features with respect to the observed source state across unobserved states. This approach uses Physics knowledge and its associated impacts on statistical and signal properties of data. We use generative learning comprising of a variational auto-encoder (VAE) based neural network model. Proposed model has a Physics influenced optimization function to achieve the data generation and adaptation of features across domains. Proposed method aims to overcome the challenges of getting the data for different progressive fault states to achieve improved predictive machine maintenance. We demonstrate the proposed method using real-world gear fault vibration time-series for synthesizing progressive faults from normal to various faulty states.