An Unsupervised TCN-based Outlier Detection for Time Series with Seasonality and Trend

Abstract

Outlier detection is challenging for time series with seasonality and trend due to the presence of local outliers. In this paper, we propose an online unsupervised deep learning based algorithm for outlier detection utilizing temporal convolutional neural network (TCN). In the proposed algorithm, firstly, the TCN network is trained using a novel loss function designed to address time series with seasonality and trend. Secondly, instead of a single global threshold for outlier detection for the entire time series, we define a set of thresholds computed based on the output of the TCN network, leading to robust detection of local outliers caused by the seasonality and the trend. The performance of the proposed algorithm is evaluated using synthetic time series. The results show that given 99% Precision, the proposed algorithm achieves at least 70% Recall and 80% F-score, which is much better than 43% Recall and 60% F-score achieved by the statistics-based seasonal extreme studentized deviate test (S-ESD) algorithm. Our algorithm also demonstrates better performance than that of the TCN based detection algorithm trained by the conventional loss function.