Title: A univariate time-series anomaly detection method for sensor network based on conditional variational AutoEncoder

Authors: Ping He; Zhifeng Wu; Sirui Hao; Hao Liu; Jiangchuan Chen; Xuefeng Song; Yipeng Liu; Junjiang He

Addresses: Information Technology Department, Sichuan Chuangang Gas Co., Ltd., Chengdu, China ' PipeChina Southwest Pipeline Co., Ltd., (Sichuan) Pilot Free Trade Zone, No. 2828, Shuzhou Road, Zhengxing Sub-district, Tianfu New Area, Chengdu City, China ' School of Cyber Science and Engineering, Sichuan University, Chengdu, China ' PipeChina Southwest Pipeline Co., Ltd., (Sichuan) Pilot Free Trade Zone, No. 2828, Shuzhou Road, Zhengxing Sub-district, Tianfu New Area, Chengdu City, China ' School of Cyber Science and Engineering, Sichuan University, Chengdu, China ' PipeChina Southwest Pipeline Co., Ltd., (Sichuan) Pilot Free Trade Zone, No. 2828, Shuzhou Road, Zhengxing Sub-district, Tianfu New Area, Chengdu City, China ' School of Cyber Science and Engineering, Sichuan University, Chengdu, China ' School of Cyber Science and Engineering, Sichuan University, Chengdu, China

Abstract: With the rapid advancement of Industry 4.0, anomaly detection secures critical infrastructure by monitoring and analysing sensor time-series data. However, existing anomaly detection methods are often suffering from incomplete temporal feature extraction and delayed detection responses. To overcome these issues, we introduce conditional variational AutoEncoder-based anomaly detection (CVAD). CVAD enhances temporal pattern learning by utilising an improved CVAE that jointly processes raw time-series data and their corresponding frequency-domain representations. To maintain robustness under concept drift, we implement a self-adjusting threshold mechanism that dynamically adjusts detection criteria, eliminating the overhead of model retraining. Experimental results on the SensorScope dataset, NASA-SMAP dataset, and NASA-MSL dataset demonstrate that the proposed model outperforms seven baseline methods in terms of the F1 score: compared with the second-best baseline TimesNet (89.45% on SensorScope, 82.99% on NASA-SMAP, 89.79% on NASA-MSL), CVAD achieves 93.35%, 93.45%, and 89.84%, respectively - representing improvements of 3.9%, 10.46%, and 0.05% in F1 score.

Keywords: sensor networks; univariate time-series data; anomaly detection; conditional variational AutoEncoder; CVAE.

DOI: 10.1504/IJSNET.2026.154339

International Journal of Sensor Networks, 2026 Vol.51 No.2, pp.115 - 128

Received: 30 Aug 2025
Accepted: 28 Oct 2025
Published online: 23 Jun 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article