Fault detection, Cyclostationary analysis, Cepstral coefficients, Cyclic CC, Cyclic spectral density, Machine learning, Random Forest, Predictive maintenance.

The growing complexity of modern industrial systems poses significant challenges for condition monitoring and early fault detection in rotating machinery. Acoustic signals generated by such equipment frequently exhibit cyclostationary properties arising from modulation phenomena associated with rotating mechanical components subject to periodic load and speed variations. Classical feature extraction techniques, such as Mel-Frequency Cepstral Coefficients (MFCCs), rely on short-time spectral representations and do not explicitly exploit cyclostationary correlations or cyclic frequencies present in the signals. This work proposes a cyclostationary generalisation of the MFCC, referred to as the Cyclic Cepstral Coefficient (CC), in which the conventional spectral energy used in the Mel filter bank is replaced by the Cyclic Spectral Density (SCD), thereby incorporating second-order information sensitive to the modulation patterns associated with mechanical faults. From the extracted coefficients a compact, physically interpretable feature vector is constructed for each sample and fed into a Random~Forest classifier trained and evaluated under different signal-to-noise ratio conditions. The approach was validated on the public MIMII dataset, which includes four types of industrial equipment: valves, pumps, fans, and slide rails, with different models for each equipment type. The results show that the proposed method achieves superior performance compared to the baseline under noisy conditions, reflecting greater sensitivity to the cyclostationary structures present in faulty operating states. The overall findings indicate that the proposed cyclostationary representation constitutes a promising alternative for fault diagnosis in noisy industrial environments.