Self-Supervised Deep Learning Framework for Anomaly Detection in Traffic Data
Clint Morris, Jidong Yang, Mi G. Chorzepa, S. Sonny Kim, Stephan A. Durham
Abstract
The current state of practice in traffic data quality control features rule-based data checking and validation processes, where the rules are subjective and insensitive to variation inherited with traffic data. In this paper, self-supervised deep learning approaches were explored to leverage the existence of multiple sources of traffic volume data, which permitted cross-checking of one data source against another for improved robustness. Two types of models were developed, aiming at detecting data anomalies at two distinct timescales. Particularly, a novel variational autoencoder (VAE)-based model was formulated for discerning data anomalies at the daily level and four recurrent model structures, including recurrent neural networks (RNN), gated recurrent units (GRU), long short-term memory (LSTM) units, and liquid time constant (LTC) networks, were evaluated for detecting anomalies in finer incremental timescales (i.e., 5-min intervals). The effectiveness of the proposed methods was demonstrated using two independent sources of traffic data from the Georgia Department of Transportation: (1) traffic counts collected by inductive loops as part of the statewide traffic count program, and (2) traffic volumes acquired by a video detection system as part of the Georgia 511, an advanced traveler information system in Georgia. Based on our experiments, the VAE-based model achieved a precision of 0.95, recall of 0.92, and F1 score of 0.94. Among the recurrent models, the fully connected LTC produced the lowest prediction error and achieved a precision of 0.82, recall of 0.88, and F1 score of 0.85.