[Paper Review] 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+2 more|arXiv (Cornell University)|Sep 14, 2023

Anomaly Detection Techniques and Applications6 citations

TL;DR

This paper proposes an automated machine learning framework that detects anomalous peak patterns in hydrological time series using synthetic data generation and hyperparameter optimization. By leveraging TimeGAN for data synthesis and optimizing among five deep learning models (TCN, InceptionTime, MiniRocket, ResNet, LSTM), the framework selects the best-performing model based on user-defined trade-offs between accuracy and computational cost, demonstrating consistent model selection on real watershed data.

ABSTRACT

This paper presents an automated machine learning framework designed to assist hydrologists in detecting anomalies in time series data generated by sensors in a research watershed in the northeastern United States critical zone. The framework specifically focuses on identifying peak-pattern anomalies, which may arise from sensor malfunctions or natural phenomena. However, the use of classification methods for anomaly detection poses challenges, such as the requirement for labeled data as ground truth and the selection of the most suitable deep learning model for the given task and dataset. To address these challenges, our framework generates labeled datasets by injecting synthetic peak patterns into synthetically generated time series data and incorporates an automated hyperparameter optimization mechanism. This mechanism generates an optimized model instance with the best architectural and training parameters from a pool of five selected models, namely Temporal Convolutional Network (TCN), InceptionTime, MiniRocket, Residual Networks (ResNet), and Long Short-Term Memory (LSTM). The selection is based on the user's preferences regarding anomaly detection accuracy and computational cost. The framework employs Time-series Generative Adversarial Networks (TimeGAN) as the synthetic dataset generator. The generated model instances are evaluated using a combination of accuracy and computational cost metrics, including training time and memory, during the anomaly detection process. Performance evaluation of the framework was conducted using a dataset from a watershed, demonstrating consistent selection of the most fitting model instance that satisfies the user's preferences.

Motivation & Objective

• To address the challenge of detecting anomalous peak patterns in time series data from a research watershed in the northeastern U.S. Critical Zone.
• To overcome the limitations of supervised anomaly detection, such as the need for labeled data and model selection complexity.
• To develop an automated pipeline that generates synthetic labeled data and optimizes deep learning models for peak anomaly detection.
• To balance detection accuracy and computational cost by selecting the most suitable model from a predefined set of five architectures.
• To evaluate the framework on real-world hydrological time series data from a critical zone observatory.

Proposed method

• Synthetic time series data are generated using TimeGAN to simulate realistic hydrological patterns.
• Artificial peak anomalies are injected into the synthetic data to create a fully labeled dataset for training.
• Five deep learning models—Temporal Convolutional Network (TCN), InceptionTime, MiniRocket, Residual Networks (ResNet), and Long Short-Term Memory (LSTM)—are evaluated for anomaly detection.
• An automated hyperparameter optimization process selects the optimal model instance based on user-defined preferences for accuracy and computational cost.
• Model performance is evaluated using metrics including detection accuracy, training time, and memory usage.
• The framework integrates synthetic data generation, model training, and selection into a single automated pipeline for hydrological applications.

Experimental results

Research questions

• RQ1 Can synthetic data generation effectively simulate realistic hydrological time series with controlled anomalous peak patterns?
• RQ2 Which deep learning architecture among TCN, InceptionTime, MiniRocket, ResNet, and LSTM performs best for detecting anomalous peak patterns under varying accuracy and computational constraints?
• RQ3 Can an automated hyperparameter optimization system consistently select the most suitable model based on user-defined trade-offs between performance and efficiency?
• RQ4 How does the framework perform on real-world time series data from a critical zone observatory?
• RQ5 To what extent does the use of synthetic data with injected anomalies improve model generalization compared to real-labeled data?

Key findings

• The framework successfully generated synthetic time series with realistic hydrological patterns and controlled anomalous peak injections using TimeGAN.
• The automated hyperparameter optimization process consistently selected the most appropriate model from the five candidates based on user-defined accuracy and computational cost preferences.
• Inference performance was validated on real watershed data, demonstrating the framework’s ability to detect anomalous peak patterns with high reliability.
• The selected model instances achieved high detection accuracy while maintaining acceptable training time and memory usage, confirming the effectiveness of the optimization strategy.
• The integration of synthetic data and automated model selection significantly reduced the dependency on costly, manually labeled real data for anomaly detection in hydrological systems.