Machine learning for anomaly detection on intelligent transportation time series data

Abstract

In intelligent transportation systems, machine learning approaches are presented to deal with time series anomaly detection. But there are always far more normal samples, making it suffer from unbalanced samples for traffic anomaly detection. In this dissertation, based on the state-of-the-art model Informer, an anomaly detection algorithm is proposed, which does not require any assumptions about the distribution of normal or anomalies. The encoder-decoder structure is applied in the generation of anomaly scores. Specifically, the encoder modified the canonical self-attention mechanism to be probability-sparse, reducing the computational complexity. The decoder is the combination of multi-attention layers and a fully connected layer to directly generate the anomaly score. Afterwards, one One-Class Support Vector Machines (OCSVM) is applied to do the classification. It has been applied in a dataset collected under real roadway circumstances and another public dataset. Experimental results have shown that the proposed algorithm performs better than several other machine learning methods.