Continuous acoustic data mining from a scattering network

In order to address the challenge of exploring new signals and recognizing sources based on both statistics and physics, a relevant time series representation tool, known as a deep scattering network, has been developed. A deep scattering network is a deep convolutional neural network that implements a cascade of convolutions with wavelet filters, a modulus function, and pooling operations. The advantage for unsupervised classification of time series is that deep scattering spectra are locally invariant to translation and preserve transient phenomena such as attack and amplitude modulation. We show that an AE adapted scattering network, combined with reduction model and clustering algorithm, is an efficient tool to perform unsupervised investigation of continuous acoustic data by applying our method on AE streaming recorded during fatigue testing: low amplitude acoustic multiplets non registered by common AE procedure has been clustered successfully and the content of the continuous acoustic background can be automatically grouped into classes of similar physical mechanisms, e.g. frictions, plasticity or electronic and mechanical noise.