Numerical simulation-based intelligent fault detection for rotary vector reducers with imbalanced classes

• A high-fidelity dynamic model of the RV reducer is developed using the lumped parameter method. • Simulation models of the RV reducer with faults are established. • Fault samples are generated using simulation models to supplement AI training data . • the performance of AI models is improved by simulation techniques under imbalance data conditions. Fault diagnosis of rotary vector reducers using artificial intelligence models plays a critical role in ensuring stable, reliable, and safe operation. However, since rotary vector reducers typically operate under normal conditions, severe class imbalance and distribution discrepancies tend to occur, which restricts the accuracy of intelligent fault diagnosis. To address the issue of severe class imbalance, an intelligent fault diagnosis method based on numerical simulation is proposed. Firstly, a lumped parameter model of the healthy rotary vector reducer is established. Subsequently, the stiffness and damping parameters are iteratively updated based on the cosine similarity between simulated and experimental vibration signals to modify the model and obtain a reliable dynamic model with improved accuracy. Secondly, mathematical models of multiple fault modes are inserted into the dynamic model to generate fault samples. The simulated fault samples are further used to supplement the training data of AI-based models. Finally, CNN, MobileNet, and Xception are selected as representative AI models, and the trained networks are subsequently used to classify unknown fault samples collected from real-world machinery, achieving classification accuracies of 98.81 %, 99.18 %, and 100 %, respectively. Experimental results indicate that the proposed method effectively addresses the issue of low classification accuracy caused by class-imbalanced fault samples in rotary vector reducers.