Figure 3

A breakthrough sequence with a total length of about 1\({\hbox { s}}\), taken from the dataset acquired during the cadaver experiments. The raw waveform was split into frames by applying a rectangular sliding window and mel spectrogram features were computed. For better illustration, the spectrograms in this Figure are plotted using a colormap, however, the features used in the implementation of this work are two-dimensional only. Furthermore, the window length is chosen arbitrarily for better visualization and is not representative for the windows which have been evaluated in this work and are much shorter. Frames (a) to (c) correspond to the non-breakthrough class, in frame (d) the breakthrough event is present and visible in the spectrogram. The features were normalized and augmented which is described in detail in the section “Pre-processing, feature extraction, data augmentation”. A modified ResNet-18⁴¹ architecture, which is introduced in the section “Deep learning model and training”, was implemented to classify breakthrough events from spectrogram features. The output dimensions of each pipeline stage are given in red color.