关键词: 可解释深度学习, 先验知识嵌入, 数字孪生, 帕累托最优化, 低延迟推理

Abstract: For the diagnosis task of key components of complex rotating machinery, the design of the diagnosis model is easily constrained by the following two characteristics：① Mechanism-driven methods are often hard to accomplish complete and accurate modelling of complex systems；② Data-driven approaches often require large-scale and high-quality data for training. To solve the above problems, a deep learning model with prior knowledge embedding is proposed to fuse prior knowledge from the mechanism and features of sensor signals. By introducing lightweight model structures, the inference delay of the model is reduced while ensuring its accuracy. Firstly, the digital twin model of key components is constructed by fusing the prior mechanism knowledge and sensor signal features. Secondly, a prior knowledge embedding module is designed to enhance the representation ability of the deep learning model for fused features. Finally, a multi-objective optimization method considering both accuracy and computing efficiency metrics is designed based on the Pareto-optimal theorem. An interpretable framework is introduced to analyse the decision-making process of the deep learning model. The results indicate that the designed digital twin of key components can provide rich prior knowledge to accelerate the convergence of the model. The training strategy based on the Pareto-optimal theorem can find the optimal solution according to settled metrics. The optimization can effectively reduce the inference latency while ensuring recognition performance.

Key words: interpretable deep learning, prior knowledge embedding, digital twin, pareto optimal, low-latency inference