An Automated Machine Learning Approach for Detecting Anomalous Peak Patterns in Time Series Data from a Research Watershed in the Northeastern United States Critical Zone

Ijaz Ul Haq, Byung Suk Lee, Donna M. Rizzo, Julia N. PerdrialarXiv 2023 · DOI · arXiv

Introduction

Environmental monitoring in critical zone watersheds relies heavily on high-frequency sensor data to track water quality and hydrological processes. However, these sensors are frequently subject to "peak-pattern" anomalies—sequences of data points that deviate from physical reality due to sensor fouling, sediment interference, or biological activity. While point outliers are relatively easy to detect, multi-point pattern anomalies present a significant challenge. Traditional supervised learning approaches to this problem are hindered by the "labeling bottleneck," where domain experts must manually annotate thousands of samples, and an "expertise gap," where hydrologists may lack the specialized machine learning knowledge required to tune complex deep learning architectures.

The HF-PPAD Framework

To address these challenges, this research introduces the Hands-Free Peak Pattern Anomaly Detection (HF-PPAD) framework. This end-to-end pipeline automates the detection of complex anomalies through a two-stage process:

1. Synthetic Data Generation via TimeGAN

The framework utilizes Generative Adversarial Networks (specifically TimeGAN) to learn the underlying distribution of a small subset of clean, real-world hydrological data. By generating a massive synthetic dataset and injecting known anomaly patterns (e.g., flat plateaus, phantom peaks, or sudden spikes), the system creates a high-fidelity, perfectly labeled training set. This approach effectively bypasses the need for manual human labeling.

2. Automated Model Selection and Optimization

HF-PPAD incorporates an Automated Machine Learning (AutoML) engine using the Optuna library. It evaluates five state-of-the-art deep learning models: InceptionTime, MiniRocket, ResNet, Temporal Convolutional Networks (TCN), and Long Short-Term Memory (LSTM) networks. The framework automatically tunes hyperparameters and allows users to weight their preferences between classification accuracy and computational efficiency, making it suitable for both high-performance servers and low-power field devices.

Experimental Setup and Datasets

The framework was validated using data from the Sleepers River Research Watershed (SRRW) in Vermont. The real-world dataset spans over six years and includes measurements for stream stage, turbidity, and Fluorescent Dissolved Organic Matter (FDOM). A synthetic benchmark of over 1 million samples was generated to train the models, which were then tested against the original sensor data.

Results and Performance

Dimensionality reduction techniques (PCA and t-SNE) confirmed that the synthetic data generated by TimeGAN maintained high fidelity to the real-world watershed distributions. All evaluated models demonstrated strong performance, with balanced accuracies ranging from 70.2% to 98.3%.

  • InceptionTime and ResNet achieved the highest accuracy but required the most computational resources.
  • MiniRocket and TCN provided the best balance of performance and efficiency.
  • LSTM was the most lightweight but showed lower accuracy compared to convolutional architectures.

Key Insights and Contextual Importance

A critical finding of this study is that models trained exclusively on synthetic data can generalize effectively to real-world sensor data. This validates the use of GANs as a proxy for manual labeling in environmental science.

The practical application of this framework is vital for accurate hydrological modeling. Cleaning sensor data is essential for analyzing Concentration-Flow (C-Q) hysteresis loops, which describe the relationship between chemical concentrations and water flow during storm events. These loops are fundamental to understanding how watersheds transport and process nutrients and pollutants.

Conceptual representation of a Concentration-Flow (C-Q) hysteresis loop, illustrating the path-dependent relationship between concentration and flow.

Furthermore, by classifying specific anomaly types (e.g., "Phantom Peaks" vs. "Flat Sinks"), HF-PPAD provides diagnostic information that can help field technicians identify specific sensor malfunctions.

Limitations and Future Directions

Currently, the framework is optimized for five specific peak-pattern types. Future work will expand the search space to include more diverse anomalies and test the system's generalizability across different geographical regions and climatic conditions, such as tropical or arid environments.

Summary updated

BibTeX
@misc{haq2023automated,
  author = {Ijaz Ul Haq and Byung Suk Lee and Donna M. Rizzo and Julia N. Perdrial},
  title = {{An Automated Machine Learning Approach for Detecting Anomalous Peak Patterns in Time Series Data from a Research Watershed in the Northeastern United States Critical Zone}},
  year = {2023},
  eprint = {2309.07992},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2309.07992},
}
Figure