Anomaly Detection in Road-Tanker Fuel Transport: A Deep Learning and Simulation Study

Fuel theft in road-tanker fleets is challenging to detect due to scarce labels, sensor noise, and distribution shifts. We present a reproducible, simulation-driven framework that combines a configurable trip simulator (geofence, speed, tank level, and outflow with realistic noise) with two detection strategies. An unsupervised LSTM Autoencoder (AE) is trained on normal windows to produce anomaly-sensitive auxiliary features. On top of these, we build: (i) an end-to-end BiLSTM+Attention model (Seq-Attn) that fuses temporal context with auxiliary features, and (ii) a hybrid LSTM+RF approach that encodes each window with an LSTM and classifies the embedding via a Random Forest. We evaluate on three scenarios---Baseline, Class Imbalance, and Distribution Shift---using trip-wise splits and theft-centric metrics (Precision, Recall, FI, PR-AUC, ROC-AUC). Results show both models achieve near-perfect detection in the baseline. Under imbalance and shift, Seq-Attn achieves consistently higher recall with almost perfect precision, while LSTM+RF remains competitive but misses slightly more thefts. These findings suggest using Seq-Attn as the primary detector, with LSTM+RF as a supplementary model for interpretability and ensemble robustness. All code and artifacts are released for end-to-end reproducibility and future validation.

Keywords:

Fuel theft detection , BiLSTM , Attention mechanism , LSTM Autoencoder , Random Forest , Simulation , Anomaly detection , Time series classification