LSTM-based approach for predicting periodic motions of an impacting system via transient dynamics

Application and comparison of feature-based classification models for multistable impact motions of percussive drilling

“…In the present study, three approaches were employed in using LSTMs for classifying the dynamical responses of the impacting system. These include (i) their direct application on the measured raw data, thereby exploiting their ability to self-extract temporal information from the data, (ii) their application on already extracted features from the raw data [10], and (iii) their application on image representations of the data. For the image data based LSTM model, the images are fed into the network as sequences of images.…”

“…By this, possible real life situation was mimicked and the need to wait for the system to reach stability before collecting data was minimised. To ensure that the used data are within the early transient periods of operation, a periods-to-stability calculation as suggested in [10] was carried out for each of the simulated time history data. During the transient periods of oscillation, the system erratically orbits around its points of attraction (Poincaré sections) in the x-v phase plane as presented in Fig.…”

“…Vibrating machinery, such as milling [1], downhole drilling [2][3][4], produce signals that are often analysed not only to ascertain their health condition, but also their performance and productivity, e.g. gears [5] and cutting [6], especially when such vibrations are normal to the machine operation [7][8][9][10]. As illustrated in Fig.…”

“…Previous studies on the resulting bit-rock impact motions have shown that some periodic motions are more efficient for the drill-bit in cutting through the rock layers compared to others [12][13][14][15]. This finding initiated the work carried out by Afebu et al [10] in which an attempt was made at developing a neural networkbased optimisation strategy for percussive drilling via multistable motion classification. A bi-directional Long-Short Term Memory (LSTM) network was used as a deep learning model to differentiate between two coexisting periodic motions.…”

“…However, due to the inhomogeneity that often characterise downhole rock layers, wider categories of impact motions are often encountered. This, thus makes it necessary to extend the introduced classification model [10] to include more categories of coexisting motions. In this present study, five categories of impact motions which were simulated from three basins of attraction using an impacting system shown in Fig.…”