Automated Identification of Critical Malfunctions of Aircraft Engines Based on Modified Wavelet Transform and Deep Neural Network Clustering

The paper considers the issues of automatic classification of vibrational states of aircraft engine malfunctions based on the use of convolutional neural network processing of vibrational measurement data presented in spectral form and the knowledge of experts with experience in interpreting spectrograms characterizing the vibrational state of aircraft engines. The developed spectrogram analysis model allows the state monitoring of aircraft engines in automatic mode both during maintenance and in flight operation. The system is able to timely notify technical personnel or crew about the appearance of signs of emergency situations, as well as the type of possible malfunctions. It is shown that the main problem affecting the quality of detection of a potential turbine malfunction is a small sample of data corresponding to malfunctioning states. It is proposed to detect emission anomalies in a small sample by recognizing a modified wavelet transform and neural network clustering, which allows more complete formation of a training sample. The data samples used in training the neural network classifier during the experimental studies were generated on the basis of existing archive files containing complete aperture data from engine vibration sensors and information about malfunctions detected in them.