隨著現地水質感測技術的發展，河川水質資料逐漸具有高時間解析度與多測站空間分布之特性，但此類高頻率資料容易受到儀器故障、生物附著、污染或外在干擾影響而產生異常值。Santos-Fernandez et al.(2025) 提出貝氏動態降秩時空模型（Bayesian Dynamical Reduced-Rank Spatio-Temporal Model, BARST），用於河川網路感測資料之異常偵測。然而，BARST 在不同測站共用同一組迴歸係數，可能限制模型對不同測站間差異之描述能力。因此，本文以 BARST 為基礎，提出具測站別迴歸係數之方法，將共同迴歸係數β擴展為測站別迴歸係數 β_s，以描述測站間之空間異質性。
本文透過模擬研究與 Herbert River 真實水質感測資料評估本文所提方法與 BARST 之異常偵測表現。模擬資料考量測站別迴歸係數與時空相依性，並加入尖峰（Spike）、高變異（High variability）、位移（Shift）與漂移（Drift）四種異常型態。結果顯示，當資料具有測站間異質性時，本文所提方法相較於 BARST 具有較佳之異常偵測表現，尤其在精確率（Precision）與 F1 分數 (F1-score)上較為明顯。真實資料分析亦顯示，本文所提方法於多數異常型態下之敏感度、精確率與 F1 分數皆略高於 BARST。
整體而言，本文所提方法在保留 BARST 的時空相依架構下，可進一步捕捉不同測站間迴歸關係差異，並改善河川網路水質感測資料之異常偵測表現。雖然真實資料之精確率與 F1 分數仍偏低，但研究結果顯示，納入測站別迴歸係數可作為處理河川網路測站異質性與提升異常偵測能力之一種可行方向，並具實際水質監測之應用價值。

This study investigates anomaly detection in high-frequency water-quality sensor data collected from stream networks. These data support environmental monitoring but are vulnerable to instrument malfunction, biofouling, calibration errors, physical obstruction, pollution, and other external disturbances. Existing Bayesian spatio-temporal methods can represent temporal dependence, stream-network spatial covariance, and predictive uncertainty, but the assumption of common regression coefficients across all monitoring stations may limit their ability to describe station-level differences. To address this issue, we extend the common coefficient vector β to station-specific coefficient vectors β_s, allowing covariate effects to vary among stations.
The proposed method is evaluated through simulation studies and real water-quality sensor data. The simulations include spikes, high variability, shifts, and drift under station-level heterogeneity. Results show that the proposed method generally improves anomaly-detection performance compared with BARST, particularly in precision and F1-score, providing a practical approach for effectively capturing spatial heterogeneity in stream-network sensor data.

Bob Carpenter, Andrew Gelman, Matthew D Hoffman, Daniel Lee, Ben Goodrich, Michael Betancourt, Marcus Brubaker, Jiqiang Guo, Peter Li, and Allen Riddell. Stan: A probabilistic programming language. Journal of Statistical Software, 76:1–32, 2017.
Varun Chandola. Anomaly detection for symbolic sequences and time series data. University of Minnesota, 2009.
Matthew D Hoffman, Andrew Gelman, et al. The no-u-turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo. J. Mach. Learn. Res., 15(1):1593–1623, 2014.
Erin E Peterson and Jay M Ver Hoef. A mixed-model moving-average approach to geostatistical modeling in stream networks. Ecology, 91(3):644–651, 2010.
Michael Rode, Andrew J Wade, Matthew J Cohen, Robert T Hensley, Michael J Bowes, James W Kirchner, George B Arhonditsis, Phil Jordan, Brian Kronvang, Sarah J Halliday, et al. Sensors in the stream: the high-frequency wave of the present, 2016.
Edgar Santos-Fernandez, Jay M Ver Hoef, Erin E Peterson, James McGree, Daniel J Isaak, and Kerrie Mengersen. Bayesian spatio-temporal models for stream networks. Computational Statistics & Data Analysis, 170:107446, 2022.
Edgar Santos-Fernandez, Jay M Ver Hoef, Erin E Peterson, James McGree, Cesar A Villa, Catherine Leigh, Ryan Turner, Cameron Roberts, and Kerrie Mengersen. Unsupervised anomaly detection in spatio-temporal stream network sensor data. Water Resources Research, 60(11):e2023WR035707, 2024.
Edgar Santos-Fernandez, Simon Denman, and Kerrie Mengersen. New Bayesian and deep learning spatio-temporal models can reveal anomalies in sensor data more effectively. Water Research, page 123928, 2025.