Change Point Detection in Time Series Data using Autoencoders with a Time-Invariant Representation

Abstract

Change point detection (CPD) aims to locate abrupt property changes in time series data. Recent CPD methods demonstrated the potential of using deep learning techniques, but often lack the ability to identify more subtle changes in the autocorrelation statistics of the signal and suffer from a high false alarm rate. To address these issues, we employ an autoencoderbased methodology with a novel loss function, through which the used autoencoders learn a partially time-invariant representation that is tailored for CPD. The result is a flexible method that allows the user to indicate whether change points should be sought in the time domain, frequency domain or both. Detectable change points include abrupt changes in the slope, mean, variance, autocorrelation function and frequency spectrum. We demonstrate that our proposed method is consistently highly competitive or superior to baseline methods on diverse simulated and real-life benchmark data sets. Finally, we mitigate the issue of false detection alarms through the use of a postprocessing procedure that combines a matched filter and a newly proposed change point score. We show that this combination drastically improves the performance of our method as well as all baseline methods.

Cadence: A Practical Time-series Partitioning Algorithm for Unlabeled IoT Sensor Streams

“…As the changes in Yahoo and Fishkiller dataset more resembles anomaly and fault detection problem, this also shows the effectiveness of RNN-based methods in anomaly data and limitation to generalize to broader CPD problems. TIRE [76] uses an autoencoder-based approach similar to Cadence to extract time and frequency domain features and finding change points using euclidean distance-based dissimilarity from the features. While it performs well on Beedance and HASC, on datasets such as Yahoo and Fishkiller that resemble anomaly data it performs worse due to its dependence on parameters based on domain knowledge.…”

“…[9,10] used a non-parametric change detection approach using Separation distance as the difference measure, but it requires pre-defined feature calculation to apply the algorithm for change-point detection. [20,51] explored change point detection using auto-encoders, but their performance for different datasets is largely influenced by window size and parameter settings. Sinn et al [82] used Maximum Mean Discrepancy (MMD) as the difference between the ordinal pattern in distribution (order structure in time series values) before and after a change-point under the assumption that the underlying time series is monotonically increasing or decreasing and for detecting multiple changepoints requires iteratively applying the method to different blocks of the time series.…”

“…LSTNet [60] uses an RNN parameterized kernel for the kernel two-sample test. TIRE [76] dissimilarity. KL-CPD [16], the current top-performing method in the literature, uses an RNN kernel learning approach by creating additional samples when sufficient data is not available.…”