Symmetric contrastive learning for robust fault detection in time-series traffic sensor data

Traffic sensor data are prone to malfunctions caused by various factors such as manufacturing defects, harsh environmental conditions, improper installation, and maintenance. While fault data detection is a well-established practice in many engineering domains, its application for quality control of traffic sensor data has not received due attention. This paper introduces a symmetric contrastive learning approach that employs a triplet network with an efficient sampling strategy, coupled with a novel cross-attention-boosted loss function for network training. To better discern the faults observed in the time-series of traffic sensor data, we employ continuous wavelet transformation (CWT) as a preprocessing step to obtain time-frequency wavelet image representation. These CWT images are used to pretrain the triplet encoder. To enhance training efficiency, we adopt an anchor-centered sampling strategy, referred to as symmetric contrastive sampling. This involves sampling a normal day of time-series traffic data as the anchor and generating three positive and three negative examples based on the domain knowledge. The loss is computed from the nine anchor-centered possible permutations of the triplet (anchor, negative, positive). This symmetric sampling approach facilitates direct contrast between positive and negative examples derived from the same anchor, leading to stronger contrastive signals for faster and more stable learning process. In comparison with traditional triplet and Siamese networks, as well as a classic threshold-based method, our proposed approach shows superior performance in detecting faulty data sequences. The experimental results demonstrate an impressive accuracy of 97.6%, precision of 97.5%, recall of 97.7%, and an F1-score of 97.6%.