Abstract
A deep neural network (DNN)-based method is proposed, which incorporates a blade-vortex interaction (BVI) aeroacoustic model and the improved Mallat-Zhong discrete wavelet transform (MZ-DWT) analysis, to detect and extract the BVI) signal. First, the optimal scale (OPS) and optimal scale vector (OPSV) features are defined based on the improved MZ-DWT to capture the dominant information of the BVI signal. Then, two types of deep neural network-based scale feature models (DNN-SFMs) are designed and trained to automatically obtain the OPS and OPSV features directly from the waveforms of the BVI signals. Finally, with the obtained OPS and OPSV features, a single-scale detector, multi-scale detector, single-scale extractor, and multi-scale extractor are derived for the BVI signal. The results of extensive experiments (BVI signals containing different types of noises are tested with each type of signal consisting of 10 000 or 9000 samples at each signal-to-noise ratio) demonstrate that the proposed detectors and extractors improve the accuracy and robustness of detection and extraction, respectively, and compared to the existing methods, the computational complexity is greatly reduced.
Highlights
Owing to its unique flight capability, the helicopter is widely used for military and civilian tasks,1 but it brings high-level noise composed of different noise types such as the blade-vortex interaction (BVI) noise, thickness noise, and loading noise
This paper focuses on the effects of the different scales on the signal features; Ws1⁄2xðnÞ is used to represent the MZ-discrete wavelet transform (DWT) of the signal xðnÞ at scale s, and Ws1⁄2xðnÞ 1⁄4 1⁄2Ws;11⁄2xðnÞ;...;Ws;n1⁄2xðnÞ;...;Ws;N1⁄2xðnÞ1ÂN, where Ws;n1⁄2xðnÞ is the Mallat-Zhong discrete wavelet transform (MZ-DWT) coefficient
To verify the performance of the proposed detectors and extractors, five types of noisecontaining BVI signals are used, as listed in Table III, where xðnÞ is the simulated noise-free BVI signal generated by Eq (1); wðnÞ is the Gaussian white noise; vðnÞ is the thickness noise collected in the anechoic chamber with different vm and l; and sðnÞ is the measured BVI signal collected in the anechoic chamber
Summary
Owing to its unique flight capability, the helicopter is widely used for military and civilian tasks, but it brings high-level noise composed of different noise types such as the blade-vortex interaction (BVI) noise, thickness noise, and loading noise. The above feature extraction methods encounter three problems: (1) the requirements of good expertise in technical fields and prior knowledge; (2) dependence on the system parameters of the helicopter (for example, in the filtering-based feature extraction method, the parameter setting of the filter depends on the frequency and energy of the main rotor harmonic); and (3) difficulty in mining new features (the traditional feature extraction methods usually depend on the existing features or evaluation criteria, which make it difficult to explore new useful features) These methods generally have such shortcomings as high computational complexity and weak robustness of the detection and extraction results. In this paper, a DNN-based method is proposed for the detection and extraction of BVI signals with the aids of a BVI aeroacoustic model and improved Mallat-Zhong discrete wavelet transform (MZ-DWT) analysis. The result is compared to the results obtained with other known methods, including the matched filter, energy detection, traditional single-scale method, traditional multi-scale method, support vector machine (SVM), PCA, and CNN
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
