A major challenge related to engine development testing is to detect abnormal observations in high-frequency sensor time series, which can hinder their convenient analysis. In practice, the use of these signals involves the calculation of spectrograms and the tracking of the vibration amplitude of different physical phenomena. The presence of these abnormal observations in the signals can go as far as to make these spectrograms completely unusable. The detection and correction of these abnormal observations become essential to continue to use the information contained in these sensors.
For proper signal analysis, an industrial anomaly detection technique has been developed by Safran Aircraft Engines. Even though this approach yields good results, it suffers from an important drawback which is the extensive intervention of an expert via an incremental fine-tuning of parameters.
To circumvent this limitation, Ezako proposed two alternative strategies, one being a variant of a classical statistical technique developed by EZAKO named Rolling Standard Deviation and the other being an architecture of the well-celebrated Deep Learning method named LSTM (Long Short Term Memory).
The study shows how our work resulted in very good detection performance while being up to 10 times faster than the expert-based method.
To download our published joint paper, please click here: Download the paper.
DETECTION OF ANOMALIES ON SIGNALS DURING AIRCRAFT ENGINE TESTS: METHODOLOGICAL COMPARISON BETWEEN HISTORICAL, STATISTICAL AND DEEP LEARNING APPROACHES. (Yann Rotrou (SAFRAN), Mourad YAHIA BACHA (SAFRAN), Julien MULLER (EZAKO), Yacine EL AMRAOUI (EZAKO))