A Transformer Neural Network For AC series arc-fault detection

Detecting series arcing faults in the electrical networks of aircraft can help mitigate dramatic consequences such as fires. Non-Artificial Intelligence algorithms often fail to generalise due to arc fault signals diversity. Most methods in the detection literature use multiple pre-processing (descriptors) associated to a machine learning model (or deep learning). These approaches require to handcraft the descriptors. We propose a deep learning approach without descriptors. We adapted a sequence-based model called a Transformer Neural Network (TNN) model to this time series problem. We repurposed the encoder of the transformer as a sequence-to-sequence model. The model takes as an input a window of electric current, with at least one period of the signals (800 Hz). The output is the label of each point in the input window. This required to propose an original manner of labelling the signals, for which we designed an automated algorithm, increasing the training supervision. Contrary to existing models on aircraft signals, our TNN model has been verified using a public experimental database of electrical-arc signals that simulates aircraft signals (230 V AC at 400−800 Hz, arcs in series with resistive loads). Our model obtained an identification accuracy of 96.3% at a 2% false positive rate. One of the significant performance of our model is that it has the lowest parameter number (2266) that can be found in scientific literature by quite some margin. TNNs are therefore an appropriate candidate for the purpose of arc fault detection, and our labelling method provides a very high temporal resolution of the output. •Automated labeling of current signals with high temporal resolution (point-by-point).•Adapting the TNN to point-by-point labeled current data using a seq2seq architecture.•The TNN’s input is a 1000-points sliding window with a sampling frequency of 100 kHz.•Detection ability with a low number of parameters that makes its implantation easier.