An EMD-based forecasting framework integrating GMM and BiLSTM for helicopter engine anomaly detection

Abstract

The safety of helicopter operations is paramount, yet early signs of potential failures often go undetected, highlighting the need for robust signal alert systems during flights. Detecting anomalies in helicopter engine behavior through vibration analysis is critically important due to the long-sequence nature and complexity of the data, which present significant challenges for real-time assessment and are not adequately addressed by traditional methods such as preset thresholds or basic statistical models, as these approaches struggle to capture intricate spatiotemporal dependencies and overlapping fault patterns in real-world scenarios. To address these challenges, we introduce a novel hybrid model that leverages Empirical Mode Decomposition (EMD) for signal decomposition and analysis, effectively overcoming the limitations of traditional approaches. EMD is particularly advantageous as it decomposes complex signals into Intrinsic Mode Functions (IMFs), enabling more accurate anomaly detection in long sequences. Following EMD, the Gaussian Mixture Model (GMM) is employed to precisely recognize various fault patterns, ensuring a robust foundation for anomaly detection. Bidirectional Long Short-Term Memory (BiLSTM) networks further enhance the model by capturing temporal dependencies in both directions, integrating critical spatiotemporal information and improving predictive accuracy. Experimental results demonstrate that this integrated EMD-GMM-BiLSTM approach is not only highly sensitive and accurate in detecting anomalies but also significantly simpler and more efficient than more complex frameworks such as encoder-decoder models or Transformers. This method ensures the operational safety of helicopters and supports the broader adoption of low-altitude economic activities by providing essential safety guarantees.