A hybrid framework for classifying and detecting aircraft cabin anomalous sounds under heavy background noise

• This paper proposes a new framework for aircraft anomalous sound detection during flight. • The combination of a classification module and an identification module is used to complete data classification and training. • Weighted Log-Mel, ArcFace classifier and a feature decomposition method are applied to improve the detection accuracy under heavy noise. • The effectiveness of the proposed method on noise resistance are verified through experiments. Anomalous sounds produced in an aircraft cabin are early warnings of potential safety hazards, yet they are hard to label and detect when masked by heavy background noise. A two-stage hybrid framework is introduced to tackle both challenges. In the first stage for anomalous sound classification, an auto-encoder extracts latent features that are clustered with a Gaussian mixture model, yielding clean and self-consistent labels for the major types of cabin anomalous sounds. And in the second stage for anomalous sound identification, a lightweight convolutional neural network, trained with ArcFace loss and aided by an anomalous feature decomposition method, detects those labeled anomalous sounds even when the signal-to-noise ratio (SNR) is very low. Evaluations on real anomalous sound recordings, synthetically mixed with cabin noise at SNRs from +8 dB to −4 dB, confirm that the framework assigns correct anomaly labels and sustains high identification accuracy of 98 % at +1 dB and 84 % at −4 dB under heavy background noise.