Anomaly Detection in Vibroacoustic Signals of a DGEN 380 Turbofan Using a LIRA Autoencoder and Gradient Boosting Framework
This paper presents a hybrid anomaly detection framework for vibroacoustic signals obtained from a DGEN 380 turbofan test bench. Traditional diagnostic approaches based on statistical thresholding exhibit limited effectiveness in non-stationary environments with multiple interacting vibration and acoustic sources. To address these challenges, a hybrid approach combining unsupervised residual modeling with supervised classification is proposed. The framework is based on the LIRA (Log-amplitude Spectral Residual Autoencoder) model, which learns a baseline representation of the system using only healthy data and identifies deviations through reconstruction-based residual analysis in both time and frequency domains. These residual representations are subsequently transformed into statistical descriptors computed across frequency sub-bands and used as input to a LightGBM classifier. The proposed approach is evaluated on a publicly available vibroacoustic turbofan dataset. The results indicate improved performance compared to classical threshold-based methods, with the boosting model achieving an F1-score of 80.60% and an AUC-ROC of 0.9025. The consistency between cross-validation and test set results further suggests good generalization capability. In addition, the integration of explainable artificial intelligence (XAI) techniques based on SHAP enables interpretation of model decisions. The analysis indicates that energy-related features, particularly RMS, play a dominant role in the detection process, while higher-order statistical descriptors contribute to robustness by capturing non-linear and impulsive behaviors. The proposed framework provides an effective and interpretable solution for anomaly detection in vibroacoustic systems and demonstrates potential for application in predictive maintenance of aeroengines and other rotating machinery.