The electricity in a wall socket is meant to be a clean wave. In India it rises and falls 50 times a second, the same way every time. Real grids are messier. A big motor starting nearby pulls the voltage down for a few cycles. Switching a capacitor bank leaves a spike that rings out. Rectifiers and chargers bend the wave out of shape. Engineers call these power-quality events.
Each kind has different causes and different fixes, so the first job is naming what happened. Monitors record far more of the wave than anyone can read by eye, which is why it helps to classify events automatically.
In 2019, at NIT Karnataka, we classified these events with long short-term memory networks (LSTMs). An LSTM reads a signal one step at a time and carries a memory forward. That lets it tell a dip lasting three cycles from a single spike, or from a slow flicker that only shows up across many cycles.
Try it below. Press a button beside the screen to inject an event, then watch a small network read the trace one cycle at a time.
Watch the notch. After the first bite the reader guesses a transient, then changes its mind as the bites keep coming. Watch the sag, too: it calls the wave normal until the dip arrives. That running memory is what a recurrent network brings to signals like these.