Detection of black box signal based on encoder-decoder fully convolutional networks

As a logger of aircraft data, the black box is the most reliable and effective means of identifying the cause of an accident after an aircraft crash. An underwater acoustic beacon was installed in the black box to deal with the black box positioning problem in the air accident at sea. The masking effect of ocean noise, coupled with the propagation loss of the ocean, causes the signal to attenuate seriously during long-distance propagation, which makes it very difficult to detect underwater signals. Inspired by the successful application of fully convolutional networks (FCN) in the field of pixel-level image classification, an encoder-decoder network with skip connnection layers, called "Unet", is proposed to enhance the underwater acoustic beacon signals represented by short-time Fourier transform (STFT) images. The experimental data show that the enhancement method based on FCN has higher signal gain than the conventional method based on adaptive line enhancer (ALE).