Machine learning based passive tap for automotive ethernet

Abstract

Modern full-duplex communication systems such as Automotive Ethernet present a significant challenge for passive network monitoring. Because both endpoints transmit simultaneously over a shared physical medium, the signal observed at any point along the cable is a superposition of two component signals that cannot be directly separated without knowledge of what each endpoint is transmitting. This thesis investigates the use of recurrent neural networks as a passive monitoring tool for full-duplex 100BASE-T1 Automotive Ethernet. A custom decoding pipeline compliant with the IEEE 802.3 standard was first developed and validated against ground-truth data, achieving a frame decode accuracy of 93.0%. This pipeline served as the primary evaluation framework for all machine learning experiments. Two passive tap configurations were investigated. In the double-tap approach, BiLSTM, LSTM, and GRU models were trained to predict component signals from two passively observed full-duplex measurements. All well-configured models exceeded the ground-truth decode accuracy baseline, with LSTM and GRU models achieving the highest accuracies of 97.0% and 96.9% respectively, despite producing higher mean absolute error than the BiLSTM. This result demonstrates that waveform MAE alone is not a sufficient indicator of passive tap performance, and that frame decode accuracy is the more meaningful evaluation metric. The single-tap investigation examined whether signal separation remains feasible from a single observation point. While more sensitive to model architecture and hyperparameter selection, the best single-tap configurations achieved mean absolute error comparable to double-tap results, suggesting that a single passive probe may be sufficient for reliable signal reconstruction with appropriate model design. Overall, these results demonstrate that machine learning is a viable approach for passive monitoring of full-duplex Automotive Ethernet, enabling frame-level decoding from simple oscilloscope probes without participation in link negotiation or access to endpoint transmission data.