Introduction
In modern industrial IoT and critical infrastructure monitoring, distinguishing between sensor noise and genuine system-wide emergencies is a significant challenge. Complex systems like power grids generate a continuous stream of co-evolving time series from hundreds of interconnected sensors. Traditional outlier detection methods often fall short by either analyzing sensors in isolation—missing "network-level" events—or relying on rigid physical models that struggle with the high-dimensional, non-stationary nature of modern grids.
This paper proposes a multi-task non-parametric framework designed to detect outliers across multiple co-evolving time series. By modeling both temporal trends and inter-series correlations, the framework identifies complex events that traditional single-series methods overlook.
Methodology
The core of the proposed approach is a Non-Parametric Multi-Task Model built on a "fitness vs. smoothness" trade-off. Unlike standard filters, this model incorporates two key dimensions:
- Temporal Smoothness: It assumes the underlying "normal" state of the system evolves smoothly over time.
- Inter-series Relatedness: This multi-task component penalizes sudden divergences between correlated sensors. If two sensors typically move in tandem, a break in that relationship is flagged as a potential outlier.
To ensure the model is distribution-agnostic, the authors utilize Bregman divergence, allowing the framework to handle various data types (e.g., Gaussian for continuous voltage, Poisson for counts). To address the computational burden of massive datasets, the authors developed the Randomized Block Coordinate Descent (RBCD) algorithm. RBCD updates the system one time-step at a time, offering superior speed and memory efficiency compared to standard optimization methods like ADMM.
Experimental Setup
The model was validated using real-world data from a power distribution system in Alameda, California. The dataset consists of measurements from five Phasor Measurement Units (PMUs), as shown in the map below.
Each PMU provided 12 channels (3-phase voltage/current magnitude and phase angle), resulting in a 60-dimensional vector per time sample. Data was collected over two months in 2015 with a 20-second resolution, and ground-truth labels were provided by power system experts.
Key Results and Discussion
The proposed method significantly outperformed traditional competitors like Hidden Markov Models (CHMM) and multivariate ARIMA (mARIMA). For network-level events, the model achieved 94.8% precision and 91.1% recall. In comparison, single-stream methods that ignore sensor relationships saw precision drop to approximately 31.8%.
Beyond accuracy, the RBCD algorithm demonstrated substantial computational advantages. As shown in the performance analysis, RBCD maintains a lower time usage as training sample sizes increase compared to other optimization strategies.
The RBCD algorithm reduced running time by at least 42.1% relative to the state-of-the-art ADMM optimizer. This efficiency allows the model to identify real-world events, such as voltage disturbances and sensor communication failures, that were previously unobservable.
Conclusion
The multi-task non-parametric framework provides a robust solution for detecting outliers in complex, co-evolving systems. By focusing on the relationships between sensors rather than just individual thresholds, it captures critical network-level events. Future work will focus on adapting the model to dynamic grid topologies where sensor correlations may change over time and scaling the approach to handle thousands of concurrent data streams.
Summary updated