可解释性智能监测诊断网络构造及航空发动机整机试车与中介轴承诊断应用

Engine health management is the key technology to improve the safety, reliability and economic affordability of aero-engine. The intelligent diagnosis method based on neural networks has achieved great success in mechanical fault diagnosis, but the current network lacks the targeted design of aero-engine due to its “black box” nature, and has not been confirmed in engineering practice. In view of these problems, this paper proposes an interpretable network construction framework for intelligent diagnosis of aero-engine and verifies it in the real engine test data. The prior information of aero-engine vibration signals is integrated into the sparse representation model, and the iterative solution algorithm of the model is unrolled to obtain an interpretable core network architecture. The interpretable sub-network via adversarial training is constructed for detection tasks, and the interpretable deep feature extraction sub-network is constructed for intelligent fault diagnosis tasks. Therefore, the network architecture proposed in this paper has a clear theoretical basis, that is, ad-hoc interpretability. In addition, a visualization method is proposed to check whether the network has learned meaningful features, making it post-hoc interpretable. The characteristics of both ad-hoc and post-hoc interpretability make the network more credible when applied to aero-engine anomaly detection and fault diagnosis. Finally, in the long-term test data analysis of a real aero-engine, the interpretable network construction proposed in this paper provides an effective and credible results for fault diagnosis of inter-shaft bearings.