Multi-Sensor Fusion-Based Anomaly Detection in Simulated CAV Environments using a D-CNN-LSTM Autoencoder

Bhukya, Balram (2025) Multi-Sensor Fusion-Based Anomaly Detection in Simulated CAV Environments using a D-CNN-LSTM Autoencoder. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (655kB) | Preview
Preview	PDF (Configuration Manual) Download (363kB) | Preview

Abstract

The proliferation of Connected and Autonomous Vehicles (CAVs) introduces significant advancements in transportation efficiency and safety, but also new vulnerabilities to sensor-based attacks and system malfunctions. Reliable anomaly detection systems are critical for ensuring the operational integrity of these complex systems. This research presents a comprehensive study on anomaly detection in a simulated CAV environment using multi-sensor data fusion. The chief idea advanced at present is the conception and experimentation with a D-CNN-LSTM autoencoder, introducing anomalous driving behaviour detection as well as understanding the complex spatio-temporal patterns pertaining to normal vehicle operation. Use of this study is made of the D2CAV dataset comprising fully synchronized data streams obtained from various vehicle sensors, each pertaining to different driving maneuvers. Furthermore, the preprocessing pipeline is extensive, comprising the following: balancing the dataset, feature engineering, and producing much of the reduction with Principal Components Analysis (PCA). Next, the performance of the D-CNN-LSTM autoencoder is evaluated using several traditional machine learning and deep learning models. These models include Random Forest, Logistic Regression, K-nearest Neighbors, Isolation Forest, a standard feed-forward Neural Network and a standalone LSTM network. Evaluation using F1-Score, Precision, Recall and AUC metrics shows the complexities in defining and detecting anomalies through multi-sensor time-series data. Although the supervised models received high F1-Scores, it was discovered that their AUC scores were low, which indicates that they usually predict the majority class. The unsupervised D-CNN-LSTM autoencoder was found to be able to learn representations from data; however, the anomaly detection effectiveness was said to be limited by the thresholding method. These results suggest the inadequacies of standard models toward this domain and thus highlight the crucial need for advanced, context-aware anomalous detection frameworks that shall ensure the safety and reliability of CAVs.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Rifai, Hicham UNSPECIFIED
Subjects:	Q Science > QA Mathematics > Computer software > Computer Security T Technology > T Technology (General) > Information Technology > Computer software > Computer Security Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	30 Jun 2026 17:27
Last Modified:	30 Jun 2026 17:27
URI:	https://norma.ncirl.ie/id/eprint/9413

Actions (login required)

View Item