Wavelet Images and Joint Deep-Machine Learning for Anomaly Detection in Airbus Helicopters Exclusively with Healthy Accelerometer Data

Abstract

Anomaly detection in aircraft systems, such as helicopters, plays a critical role in ensuring operational safety and reliability. A primary challenge in this domain is the unavailability of sufficient labeled data, with the majority of available datasets comprising only healthy conditions. This paper addresses these challenges by proposing a robust pipeline for anomaly detection using accelerometer data collected from Airbus helicopter systems. Specifically, the accelerometer data, which often exhibits complex spatiotemporal characteristics, is first transformed into images using Continuous Wavelet Transform (CWT). Principal Component Analysis (PCA) is then applied to reduce the dimensionality of the data while retaining the maximum variance (i.e., $100 \%$), thereby compressing it with minimal loss. Subsequently, a Fast AutoEncoder (FAE), trained using a pseudo-inverse method on healthy data, is employed to predict anomalies. The prediction errors are analyzed through several small-scale machine learning models, including Gaussian Mixture Models (GMM), Support Vector Machines (SVM), and AutoEncoders trained with Stochastic Gradient Descent (AESGD). To optimize performance, Bayesian optimization is integrated for hyperparameter tuning, including threshold-based classification. Key data visualizations are presented to demonstrate the effectiveness of CWT transformation, the quality of the data processed by FAE, and the potential for separating data patterns. Finally, performance comparisons based on ROC, AUC, and classification threshold demonstrate the efficacy of the proposed pipeline, with SVM achieving an AUC of 0.83281, which is $55.46 \%$ higher than GMM and 7.28% higher than AESGD. This highlights its superior ability to separate anomalous patterns compared to other methods.