Collective anomaly detection in time series using pitch frequency and dissimilarity features

Date

2023-6-12

Author

Erkuş, Ekin Can

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

87
views

0
downloads

Cite This

Collective anomalies appear in the majority of time series data modalities due to a variety of factors. They appear frequently in biomedical signals as a result of electrode displacement, motion, or faulty equipment. These anomalies have a negative impact on model and analysis performance and are frequently identified in order to be eliminated or detected in order to observe unwanted data behavior. This thesis describes a novel method for detecting collective anomalies in quasi-periodic time series data. By leveraging pitch frequency estimation techniques commonly used in audio signal processing, the proposed algorithm combines the strengths of both the time and frequency domains. It provides a comprehensive view of anomalous patterns and can be customized and adapted to different domains and datasets, making it useful for a wide range of applications. By employing a sliding windows approach and utilizing previous data information to dynamically learn structural patterns, the proposed algorithm also excels in real-time anomaly detection. It is effective in detecting subject-specific anomalies, although it may not locate single-sample outliers that do not significantly affect window properties. The algorithm was developed specifically for quasi-periodic data and may be limited in its applicability to non-quasi-periodic time series data. Both synthetically generated and benchmark electrocardiogram (ECG) datasets are used to assess the effectiveness of the proposed algorithm under a variety of conditions. The performance of the proposed approach is compared to other features commonly used in anomaly detection, as well as some benchmark time series anomaly detection algorithms. The findings show that the proposed method consistently outperforms the compared algorithms in detecting both outlier-like and inlier-like anomalies. It also outperforms other non-parametric approaches in terms of computational efficiency.

Subject Keywords

anomaly detection, time series, pitch frequency, dissimilarity, sliding windows

URI

https://hdl.handle.net/11511/104738

Collections

Graduate School of Natural and Applied Sciences, Thesis