Open cross-domain data fusion for fault diagnosis of complex equipment

Abstract

Abstract In order to address the technical challenge of acquiring a significant number of fault samples from actual industrial sites, this paper proposes a fault diagnosis method that utilizes a deep learning model driven by cross-domain data fusion. Firstly, a high-fidelity digital twin model of a planetary gearbox fault diagnosis test platform is constructed, enabling the acquisition of simulated vibration signals from the real-time speed drive model. Secondly, empirical mode decomposition is applied to both the simulated fault signal and the measured normal signal. Three IMF components with high variance contribution rates of the measured normal signal are screened and reconstructed with the IMF component of the simulated fault signal to generate the fused signal. Then, a deep residual network model based on the channel attention mechanism is constructed. Finally, the network models are trained and tested using the acquired data. Evaluation indices, such as recall and accuracy, are employed to evaluate the engineering reliability of the fused data. The experimental results demonstrate that the proposed cross-domain data fusion-driven method offers a novel approach to address the technical challenge of limited fault samples in actual industrial sites.