Abstract: In-situ inspection of aero-engines is a key technology for identifying potential early faults and ensuring operational safety under the current scheduled maintenance regime. It also serves as a crucial support mechanism for achieving precise prediction and optimizing maintenance decisions within the future condition-based maintenance framework. While AI-enabled in-situ inspection has recently garnered widespread attention from both academia and industry, the “black-box” nature of deep learning models has created a dilemma in aviation applications where operators are “willing but hesitant to deploy.” Addressing the urgent demand for high-precision detection of Intake plates cracks, this paper proposes an interpretable intelligent acoustic signature diagnosis method. By fusing prior knowledge of the modal response of strut crack acoustic signatures, a multi-scale sparse representation model is constructed. The optimization process is then mapped into a network architecture, ensuring the network construction is theoretically grounded. Furthermore, an anomaly detection framework based on adversarial training is introduced, combined with multi-scale feature reconstruction to conduct post-hoc interpretable analysis, making diagnostic conclusions traceable. Based on this, a time-frequency enhancement method is proposed to extract sensitive acoustic signature parameters characterizing cracks, realizing a “secondary verification” of the network outputs. This jointly addresses the coexistence of false alarms and missed detections in engineering applications. Finally, field measurement data from the in-situ inspection of a certain type of aero-engine inlet strut validates the effectiveness and engineering applicability of the proposed method, providing technical support for high-precision in-situ detection of intake plates cracks.

Key words: aero-engine, intake plates crack, acoustic signature diagnosis, interpretable anomaly detection, sparse unrolling network, time-frequency enhancement