Abstract:
The evolution of technologies in the automobile industry has increased the complexity of
the in-vehicle electrical network. This poses a new challenge for testers to do troubleshooting
work in massive log files of CAN (Controller Area Network) data obtained from cars. The
Controller Area Network (CAN) bus is one of the in-vehicle communication networks that
enable each ECU to communicate with all other ECUs in a system. Data from these com ponents are required to be monitored to detect any anomalies or faults in the system during
testing. Manual testing for detecting anomalies has limitations such as - high cost and high
time consumption. To overcome these limitations AI-based anomaly detection techniques
can be employed. This project deals with an anomaly detection mechanism in which vari ous anomalies (manually introduced) in the captured CAN data logs are identified using an
LSTM autoencoder. The detected anomalies can further be classified using BiLSTM clas sifier. The normal CAN data logs from the HIL simulator are used for training the LSTM
AE model. Optuna framework is employed for hyperparameter tuning. The model which
is trained using the optimal hyperparameters is tested using anomaly-introduced data and
finally performing dynamic thresholding on the reconstruction error obtained from the origi nal and the predicted values of the model, anomalies can be identified. The anomaly-injected
CAN logs are used to train the BiLSTM model, which can classify the anomaly type from
the anomaly-detected data. Experimental results have shown that the LSTM autoencoder
model obtained an accuracy of 99%, precision of 81.22%, recall of 98.2%, and F1-Score of
88%, and the BiLSTM classifier can classify the anomalies with an accuracy of 99%
